(19)
(11)EP 3 579 196 A1

(12)EUROPEAN PATENT APPLICATION

(43)Date of publication:
11.12.2019 Bulletin 2019/50

(21)Application number: 18176132.1

(22)Date of filing:  05.06.2018
(51)International Patent Classification (IPC): 
G06T 19/00(2011.01)
G06T 17/20(2006.01)
(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR
Designated Extension States:
BA ME
Designated Validation States:
KH MA MD TN

(71)Applicants:
  • Sminchisescu, Cristian
    Bucharest (RO)
  • Zanfir, Mihai
    Bucharest (RO)
  • Popa, Alin-Ionut
    Sat Netgoiesti (com. Brazi), Jud. Prahova (RO)
  • Zanfir, Andrei
    Mun. Bucuresti sect 4 (RO)
  • Marinoiu, Elisabeta
    Ploiesti, jud Prahova (RO)

(72)Inventors:
  • Sminchisescu, Cristian
    Bucharest (RO)
  • Zanfir, Mihai
    Bucharest (RO)
  • Popa, Alin-Ionut
    Sat Netgoiesti (com. Brazi), Jud. Prahova (RO)
  • Zanfir, Andrei
    Mun. Bucuresti sect 4 (RO)
  • Marinoiu, Elisabeta
    Ploiesti, jud Prahova (RO)

(74)Representative: Betten & Resch 
Patent- und Rechtsanwälte PartGmbB Maximiliansplatz 14
80333 München
80333 München (DE)

  


(54)HUMAN CLOTHING TRANSFER METHOD, SYSTEM AND DEVICE


(57) There is provided a method and system comprising determining in a first image a person and a first 3D human pose and body shape fitting model, wherein the person has a first pose and a first clothing, determining in a second image a person and a second 3D human pose and body shape fitting model, wherein the person has a second pose and a second clothing, and generating, by use of the first 3D human pose and body shape fitting model and by use of the second 3D human pose and body shape fitting model, an image comprising the person of the first image in the first pose and with the second clothing and/or generating, by use of the first 3D human pose and body shape fitting model and by use of the second 3D human pose and body shape fitting model, an image comprising the person of the second image in the second pose and with the first clothing.




Description

TECHNICAL FIELD



[0001] The present invention is directed to image processing method, system and device. Particularly, the present invention is directed to generating an image in view of two already present images. Each one of the two already present images comprise image data of at least one person. An image is generated that synthesizes the pose or the at least one person in one image and the clothing of the at least one person in the other image.

BACKGROUND



[0002] People are of central interest in images and video, so understanding and capturing their pose and appearance from visual data is critically important. While problems like detection or 2d pose estimation have received considerable attention and witnessed significant progress recently, appearance modeling has been less explored comparatively, especially for bodies and clothing, in contrast to faces. One setback is that people are extremely sensitive to invalid human appearance variations and immediately spot them. This is to a large extent true for faces, as people are sharply tuned to fine social signals expressed as subtle facial expressions, but also stands true for human body poses, shapes and clothing. This makes it difficult to capture and possibly re-synthesize human appearance in ways that pass the high bar of human perception. While the realistic 3d human shape and appearance generation, including clothing, has been a long standing goal in computer graphics, with impressive studio results that occasionally pass the Turing test, these usually require complex models with sophisticated layering, manual interaction, and many cameras, which makes them difficult to use at large scale. For this purpose, flexible methods that can be learned from data and can synthesize realistic human appearance are of obvious value. Arguably, even more important would be methods that can be controlled by image evidence in some way. For instance one may not just aim to generate plausible human shape and appearance in isolation - hard as this may be - but also condition on specific elements of pose and appearance in a given image in order to synthesize new ones based on it.

SUMMARY



[0003] Herein, we formulate a new problem called human appearance transfer. Given a single source and a single target image of a person, each with different appearance, possibly different body shape and pose, the goal is to transfer the appearance of the person in the first image into the one of the person of the target image while preserving the target clothing and body layout. The problem is challenging as people are in different poses and may have different body shapes. A purely image warping or image to image translation approach would not easily generalize due to the large number of degrees of freedom involved in the transformation, e.g. the effect of articulation, depth and body shape on appearance. We provide a first solution that relies on fitting state-of-the-art 3d human pose and body models to both the source and the target images, transferring appearance using barycentric methods for commonly visible vertices, and learning to color the remaining ones using deep image synthesis techniques with appropriately structured 2d and 3d inputs. Example images, perceptual user studies, Inception scores (see T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, X. Chen, and X. Chen: "Improved techniques for training gans", in NIPS, 2016, wherein disclosure of said publication is incorporated herein by reference), and the response of a state-of-the-art person detector confirm that the generated images of humans are perceptually plausible.

[0004] See hereto Fig. 1, which shows a bidirectional transfer automatically produced by our method. We transfer from one image (first column) to the second (second column) and vice-versa, with automatically generated images shown on the third and fourth columns.

[0005] We propose an automatic person-to-person appearance transfer model based on explicit parametric 3D human representations and learned, constrained deep translation network architectures for photographic image synthesis. Given a single source image and a single target image, each corresponding to different human subjects, wearing different clothing and in different poses, our goal is to photorealistically transfer the appearance from the source image onto the target image while preserving the target shape and clothing segmentation layout. Our solution to this new problem is formulated in terms of a computational pipeline that combines (1) 3D human pose and body shape estimation from monocular images, (2) identifying 3d surface colors elements (mesh triangles) visible in both images, that can be transferred directly using barycentric procedures, and (3) predicting surface appearance missing in the first image but visible in the second one using deep learning-based image synthesis techniques. Our model achieves promising results as supported by a perceptual user study where the participants rated around 65% of our results as good, very good or perfect, as well in automated tests (Inception scores and a Faster-RCNN human detector responding very similarly to real and model generated images). We further show how the proposed architecture can be profiled to automatically generate images of a person dressed with different clothing transferred from a person in another image, opening paths for applications in entertainment and photo-editing (e.g. embodying and posing as friends or famous actors), the fashion industry, or affordable online shopping of clothing.

[0006] The objects of the present application are solved by subject matters of independent claims. Further exemplary embodiments are provided in independent claims.

[0007] According to an aspect of the present invention a method is provided that comprises:

determining in a first image a person and a first 3D human pose and body shape fitting model,

wherein the person has a first pose and a first clothing; determining in a second image a person and a second 3D human pose and body shape fitting model, wherein the person has a second pose and a second clothing; and generating, by use of the first 3D human pose and body shape fitting model and by use of the second 3D human pose and body shape fitting model, an image comprising the person of the first image in the first pose and with the second clothing and/or

generating, by use of the first 3D human pose and body shape fitting model and by use of the second 3D human pose and body shape fitting model, an image comprising the person of the second image in the second pose and with the first clothing.



[0008] According to an embodiment, the first 3D human pose and body shape fitting model comprises a first number of vertices and a first set of faces that form a first triangle mesh, and the second 3D human pose and body shape fitting model comprises a second number of vertices and a second set of faces that form a second triangle mesh.

[0009] According to an embodiment, the first 3D human pose and body shape fitting model represents a tuple St = (Ct, Dt, Mt), wherein, with regard to the first triangle mesh, Ct encodes a RGB color at each vertex position, Dt encodes a disparity 3d depth map of the first triangle mesh with respect to the first image, and Mt encodes a visibility of each vertex. According to an embodiment, the second 3D human pose and body shape fitting model represents a tuple Ss = (Cs, Ds, Ms), wherein, with regard to the second triangle mesh, Cs encodes a RGB color at each vertex position, Ds encodes a disparity map of the first triangle mesh with respect to the first image, and Ms encodes a visibility of each vertex.

[0010] According to an embodiment, the first 3D human pose and body shape fitting model is determined or obtained by determining, with regard to the person in the first image and by using a multitask deep neural network model, corresponding 2D body joint locations, semantic body part segments and 3D pose and by refining the first 3D human pose and body shape fitting model by executing non-linear optimization. According to an embodiment, the second 3D human pose and body shape fitting module is determined or obtained by determining, with regard to the person in the second image and by using the multitask deep neural network model, corresponding 2D body joint locations, semantic body part segments and 3D pose and by refining the second 3D human pose and body shape fitting model by executing non-linear optimization.

[0011] According to an embodiment, the refining the first 3D human pose and body shape fitting model by executing non-linear optimization comprises aligning, with regard to the person in the first image, a corresponding articulated human body mesh with a semantic segmentation layout obtained by using the multitask deep neural network model with regard to the person in the first image. According to an embodiment, the refining the second 3D human pose and body shape fitting model by executing non-linear optimization comprises aligning, with regard to the person in the second image, a corresponding articulated human body mesh with a semantic segmentation layout obtained by using the multitask deep neural network model with regard to the person in the second image.

[0012] According to an embodiment, the non-linear optimization is performed by executing semantic fitting where vertices of the respective mesh model carry a body type which is matched to the corresponding body type pixels detected in the image in the sense that vertices of a particular type of 3d model should onto pixels detected of the same type in the image.

[0013] According to an embodiment, the generating of the image comprises: determining a common set of visible vertices visible in the first triangle mesh with regard to the person in the first image and in the second triangle mesh with regard to the person in the second image; determining a divergent set of vertices visible in the first triangle mesh with regard to the person in the first image and not visible in the second triangle mesh with regard to the person in the second image; assigning colors to the vertices in the common set of visible vertices by using a barycentric transfer; training a first neural network for body color completion for assigning colors to vertices of the first triangle mesh that are visible and that are not present in the common set of visible vertices; and training a second neural network for human appearance synthesis for assigning colors of resulting appearance image of the first person, that are part of clothing, but cannot be modeled by the mesh model of the first person and the first neural network for body color completion.

[0014] According to an embodiment, training a first neural network comprises using weakly-supervised learning based on sampling subsets of visible vertices of a 3d mesh fitted to images of people, as source and target set for self-training, respectively.

[0015] According to an embodiment, the first neural network and the second neural network are first and second neural network modules, respectively. According to an embodiment, the parameters of all neural network modules, i.e. for example of the first and the second module, are trained in an end-to-end process and using a multi-task loss. According to an embodiment, the neural network modules are part of a larger integrated system that is trained in a consistent manner.

[0016] According to an embodiment, the second neural network for human appearance synthesis is trained to predict its final output, given among inputs, clothing segmentation layout, disparity map and results predicted by the first neural network for body color completion.

[0017] According to an embodiment, the generating of the image comprises transferring a clothing layout of the second image to a clothing layout of the first image.

[0018] According to an aspect of the present invention, a device is provided that configured to execute steps of a method as outlined above and as described in more detail below. The device comprises according to an embodiment one or more processing means (e.g., processors) that are configured to correspondingly execute said steps.

[0019] According to an aspect of the present invention, a system is provided that configured to execute steps of a method as outlined above and as described in more detail below. According to an embodiment, the system comprises one or more components or modules respectively that are configured to correspondingly execute said steps.

DETAILED DESCRIPTION



[0020] Hereinafter, some exemplary embodiments will be described in detail with reference to the accompanying drawings. Regarding the reference numerals assigned to the elements in the drawings, it should be noted that the same elements will be designated by the same reference numerals, wherever possible, even though they are shown in different drawings. Also, in the description of embodiments, detailed description of well-known related structures or functions will be omitted when it is deemed that such description will cause ambiguous interpretation of the present disclosure. Generally, it has to be noted that all arrangements, devices, systems, modules, components, units, and means and so forth described herein could be implemented by software or hardware elements or any kind of combination thereof. All steps which are performed by the various entities described in the present application as well as the functionality described to be performed by the various entities are intended to mean that the respective entity is adapted to or configured to perform the respective steps and functionalities. Further, the method of the present invention and its various steps are embodied in the functionalities of the various described apparatus elements. Moreover, any of the embodiments and features of any of the embodiments, described herein, may be combined with each other, i.e. may supplement each other, unless a combination is explicitly excluded.

[0021] Our work relies on 2d human detection and body part labeling (see F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it smpl: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016; A. Popa, M. Zanfir, and C. Sminchisescu: "Deep Multitask Architecture for Integrated 2D and 3D Human Sensing", in CVPR, July 2017; and K. He, G. Gkioxari, P. Doll'ar, and R. Girshick: "Mask r-cnn", in ICCV, 2017; wherein the disclosures of said documents are included herein by reference), 3d pose estimation (see A. Popa, M. Zanfir, and C. Sminchisescu: "Deep Multitask Architecture for Integrated 2D and 3D Human Sensing", in CVPR, July 2017; F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it smpl: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016; A. Toshev and C. Szegedy. Deeppose: "Human pose estimation via deep neural networks", in CVPR, 2014; wherein the disclosures of said documents are included herein by reference), parametric 3d human shape modeling (see R. Goldenthal, D. Harmon, R. Fattal, M. Bercovier, and E. Grinspun: "Efficient simulation of inextensible cloth", ACM Transactions on Graphics (TOG), 26(3):49, 2007; F. Xu, Y. Liu, C. Stoll, J. Tompkin, G. Bharaj, Q. Dai, H.-P. Seidel, J. Kautz, and C. Theobalt: "Video-based characters: Creating new human performances from a multi-view video database", in ACM SIGGRAPH 2011 Papers, SIGGRAPH, pages 32:1-32:10, New York, NY, USA, 2011, ACM; R. Narain, A. Samii, and J. F. O'Brien: "Adaptive anisotropic remeshing for cloth simulation", ACM transactions on graphics (TOG), 31(6):152, 2012; M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. Smpl: A skinned multi-person linear model, SIGGRAPH, 34(6):248, 2015; wherein the disclosures of said documents are included herein by reference) procedures devoted to the semantic segmentation of clothing (see E. Simo-Serra, S. Fidler, F. Moreno-Noguer, and R. Urtasun: "A high performance crf model for clothes parsing", in ACCV, 2014; K. Gong, X. Liang, X. Shen, and L. Lin "Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing", in CVPR, July 2017; Z. Liu, P. Luo, S. Qiu, X. Wang, and X. Tang: "Deepfashion: Powering robust clothes recognition and retrieval with rich annotations" in CVPR, pages 1096-1104, 2016; M. Hadi Kiapour, X. Han, S. Lazebnik, A. C. Berg, and T. L. Berg: "Where to buy it: Matching street clothing photos in online shops"; in ICCV, December 2015; wherein the disclosures of said documents are included herein by reference), as well as image translation and synthesis methods (see P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros: "Image-to-image translation with conditional adversarial networks", in CVPR, 2017; Q. Chen and V Koltun: "Photographic image synthesis with cascaded refinement networks", in ICCV, October 2017; A. Nguyen, J. Yosinski, Y. Bengio, A. Dosovitskiy, and J. Clune: "Plug & play generative networks: Conditional iterative generation of images in latent space", in CVPR, 2017; C. Ledig, L. Theis, F. Husz'ar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, et al.: "Photo-realistic single image super-resolution using a generative adversarial network", in CVPR, 2017; J.-Y Zhu, T. Park, P. Isola, and A. A. Efros: "Unpaired imageto-image translation using cycle-consistent adversarial networks", in ICCV, 2017; Q. Chen and V. Koltun: "Photographic image synthesis with cascaded refinement networks", in ICCV, October 2017; C. Yang, X. Lu, Z. Lin, E. Shechtman, O. Wang, and H. Li: "High-resolution image inpainting using multi-scale neural patch synthesis", in CVPR, 2017; wherein the disclosures of said documents are included herein by reference).

[0022] Modeling the human appearance is a vast topic that has been approached on several fronts. One is through modifications of real images (see C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu: "Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments", PAMI, 2014; L. Pishchulin, A. Jain, M. Andriluka, T. Thormaehlen, and B. Schiele: "Articulated people detection and pose estimation: Reshaping the future", in CVPR, June 2012; wherein the disclosures of said documents are included herein by reference), although the results are not entirely realistic. Computer graphics pipelines are also used, either in a mixed reality setting - where a moderately realistic graphics model is rendered in a real scene in a geometrically correct manner (see C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu: "Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments", PAMI, 2014; wherein the disclosures of said document is included herein by reference) - or e.g. by fitting a SCAPE model to real images (see W. Chen, H.Wang, Y. Li, H. Su, Z.Wang, C. Tu, D. Lischinski, D. Cohen-Or, and B. Chen: "Synthesizing training images for boosting human 3d pose estimation", in 3D Vision (3DV), 2016 Fourth International Conference on, pages 479-488, IEEE, 2016; wherein the disclosures of said document is included herein by reference). In the former, the graphics character is still not photo-realistic; in the latter, clothing geometry is lacking. Detailed, accurate human shape estimation from clothed 3d scan sequences (see C. Zhang, S. Pujades, M. Black, and G. Pons-Moll: "Detailed, accurate, human shape estimation from clothed 3d scan sequences", in CVPR, 2017; wherein the disclosures of said document is included herein by reference) can produce very good results but the acquisition setup is considerably more involved. Models to realistically capture complex human appearance including clothing in a laboratory setup, based on multiple cameras and relatively simple backgrounds appear in V. Leroy, J.-S. Franco, and E. Boyer: "Multi-view dynamic shape refinement using local temporal integration", in ICCV, 2017, wherein the disclosures of said document is included herein by reference. Procedures directed to the realistic acquisition of clothing exist (see G. Pons-Moll, S. Pujades, S. Hu, and M. Black. ClothCap: "Seamless 4D clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH), 36(4), 2017, two first authors contributed equally; C. Zhang, S. Pujades, M. Black, and G. Pons-Moll: "Detailed, accurate, human shape estimation from clothed 3d scan sequences", in CVPR, 2017; wherein the disclosures of said documents are included herein by reference), but rely on an existing set of 3d models of garments and a 3d scanning device.

[0023] The methodology reviewed in the previous paragraph achieves excellent results under the application constraints it was designed for. However, some requires manual interaction, multiple cameras, simple backgrounds, specialized scanners, or complex modeling setups. In contrast, we aim at automatic appearance modeling in situations where one has no control on the acquisition setup and is given a minimal number of images (one or two). The idea is to exploit precise, but inherently limited in coverage, geometric estimation methods for the human pose and shape, and complement them with learning techniques, in order to achieve photo-realistic appearance transfer for specific images. There is relatively little research focusing on human appearance generation based on a combination of geometric and learning methods. One notable exception is the recent work by C. Lassner, G. Pons-Moll, and P. V. Gehler: "A generative model of people in clothing", in ICCV, 2017 (wherein the disclosure of said document is included herein by reference) which is able to generate realistic images of people given their silhouette, pose and clothing segmentation. The method relies on a variational auto-encoder (see T. D. Kulkarni, W. Whitney, P. Kohli, and J. B. Tenenbaum: "Deep convolutional inverse graphics network", in NIPS, 2015; wherein the disclosure of said document is included herein by reference) and a GAN (see I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D.Warde-Farley, S. Ozair, A. Courville, and Y. Bengio: "Generative adversarial nets", in NIPS, 2014; P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros: "Image-to-image translation with conditional adversarial networks", in CVPR, 2017; wherein disclosures of said documents are included herein by reference) for realistic image generation. However, it is not obvious how this model would perform inference for the appearance given an image, or how can it condition on a particular appearance and photo-realistically transfer it to a new pose. The human appearance transfer between two monocular images falls out of the domain of applicability of models like C. Lassner, G. Pons-Moll, and P. V. Gehler: "A generative model of people in clothing", in ICCV, 2017 (wherein the disclosure of said document is included herein by reference), and is the new problem defined and confronted in this research.

HUMAN APPEARANCE TRANSFER



[0024] Given a pair of RGB images - source and target, denoted by Is and It, each containing a person -, the main objective of our work is to transfer the appearance of the person from Is into the body configuration of the person from It, resulting in a new image Ist (the procedure is symmetric, as we can transfer in both directions.). Our proposed pipeline is shown in Fig. 2 and details are given in the next sections.

[0025] Fig. 2 visualizes a human appearance transfer pipeline and particularly the steps executed according to an embodiment of the present invention. Given only a single source and a single target image, each containing a person, with different appearance and clothing, and in different poses, our goal is to photo-realistically transfer the appearance from the source image onto the target image while preserving the target shape and clothing segmentation layout. The problem is formulated in terms of a computational pipeline that combines (i) 3d human pose and body shape fitting from monocular images shown in Fig. 3, together with (ii) identifying 3d surface colors corresponding to mesh triangles visible in both images, that can be transferred directly using barycentric procedures, (iii) predicting surface appearance missing in the target image but visible in the source one using deep learning image synthesis techniques - these will be combined using the Body Color Completion Module detailed in Fig. 5. The last step, (iv), takes the previous output together with the clothing layout of the source image warped on the target image (Clothing Layout Warping) and synthesizes the final output. If the clothing source layout is similar to the target, we bypass the warping step and use the target clothing layout instead.

3D Human Pose and Body Shape Fitting



[0026] Human Detection & Body Parts Segmentation. To detect each person and infer critical semantic and geometric information, each image is fed through the Deep Multitask Human Sensing (DMHS) network (see, e.g., A. Popa, M. Zanfir, and C. Sminchisescu: "Deep Multitask Architecture for Integrated 2D and 3D Human Sensing", in CVPR, July 2017; wherein the disclosure of said document is incorporated herein by reference), a state-of-theart predictor for body part labeling (semantic segmentation) and 3d pose estimation. DMHS is a multi-stage architecture, in which each stage refines the output from the previous stage, producing a tuple (J;B;R), where

is the set of 2d body joint configurations, B

is the body part labeling map, and

is the 3d body joint configuration of the person detected in an image.

[0027] 3d Body Shape Fitting. We use the prediction of DMHS with the fitting method of A. Zanfir, E. Marinoiu, and C. Sminchisescu. Monocular: "3D Pose and Shape Estimation of Multiple People in Natural Scenes - The Importance of Multiple Scene Constraints", in CVPR, 2018 (the disclosure of said document is included herein by reference) in order to estimate the human 3d body shape and pose from an image. The representation is based on the 3d SMPL body model (see M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. Smpl: "A skinned multi-person linear model", SIGGRAPH, 34(6):248, 2015; wherein the disclosure of said document is incorporated herein by reference). A commonly used pipeline for fitting the model (see F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it smpl: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016; wherein the disclosure of said document is incorporated herein by reference) relies on minimizing a sparse cost error over detected 2d human joint locations. However, the method of A. Zanfir, E. Marinoiu, and C. Sminchisescu: "Monocular 3D Pose and Shape Estimation of Multiple People in Natural Scenes - The Importance of Multiple Scene Constraints", in CVPR, 2018 (wherein the disclosure of said document is incorporated herein by reference) utilizes all information available in the tuple (J;B;R) and the image. The 3d pose is initialized using R and refined so that each body part of the model aligns with the corresponding semantic segmentation labels in the image, based on DMHS estimates J and B (see Fig. 3). Please notice the difference between our body shape fitting procedure and the one of F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it smpl: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016, illustrated in Fig. 4. The head and arm orientations of our estimate are closer to the perceived one, due to a superior DMHS initialization (as opposed to a canonical T-pose) and the use of dense body parts semantic image segmentation labels during fitting.

[0028] The estimated 3d body model consists of a fixed number of Nv = 6890 vertices and a set of Nf = 13776 faces,

that forms the triangle mesh. We define the fitted 3d body shape model as a tuple S = (C;D;M), where

encodes the RGB color at each vertex position,

encodes the disparity map of the fitted 3d mesh with respect to the image and M ∈ {0,1}Nv encodes the visibility of each vertex. Given the source and target images Is and It, we obtain human pose and shape estimates as mesh structures Ss and St, respectively.

[0029] Fig. 4 shows a comparison of 3d human pose and shape estimation results of our employed method (green), that combines the 2d and 3d predictions of a deep multitask neural network with pose refinement, with F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it smpl: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016 (red). As our model, illustrated in Fig. 3, is initialized using an image-sensitive 3d pose predictor (as opposed to fixed initialization) and is fitted to the full semantic person layout (the complete semantic segmentation of the person into body parts, as opposed to only a sparse set of human body joints), it tends to better cope with the angle and the orientation of the head and arms, as well as the body proportions.

Body Color Completion



[0030] Fig. 5 shows a Body Color Completion module or process. Given a Source (S) and a Target (T) image, with 3d human pose and shape fitted automatically and represented using 3d meshes, we compute two sets of mesh vertices: ones that are visible in both images (T ∩ S), and ones that are only visible in the target T\S. We propagate (copy) the color of the intersection set using barycentric transfer and predict the difference set using a vertex color completion network. The network is trained to predict a subset of vertex colors from other subsets of vertex colors. Training data for this network can be readily available by sampling any 'cut' through the set of visible vertices obtained by fitting a 3d model to a single image. However, in practice we make training more typical by using targets that come from S\T in pairs of images.

[0031] We are first interested in estimating the pixel colors for the projected visible surface of St, denoted as Is→t, using the pixel colors on the projected visible surface of Ss.

[0032] Barycentric transfer. We begin by defining the common set of visible vertices ΛsΛt = {i|Mt(i) = 1 Λ Ms(i) = 0,1 ≤ iNv} and select the corresponding mesh faces FsΛt). For each face fFsΛt), we project it on Is and Is→t. For each pixel location in the projection off on Is, we find its corresponding pixel location in the projection on Is→t using barycentric triangle coordinates. Finally, we copy the color information from one location to another.

[0033] Vertex Color Completion Network. The remaining set of visible vertices in St, Λs\t = {i|Mt(i) = 1 Λ Ms(i) = 0,1 ≤ i ≤ Nv} needs to be colored. We rely on learning the implicit correlations among various body parts in order to propagate appearance information from the already colored vertex set Cs to ΛsΛt. Such correlations, effectively forms of pattern completion, are learned automatically from training data using a neural network.

[0034] Learning for Mesh Color Completion. We are given as inputs the color set Cs, the visibility mask Ms ∈ {0,1}Nv×Nv and a binary mask that encodes the vertices we wish to color MΛt\s ∈ {0,1}Nv×Nv, i.e. visibility values are replicated along columns. The output is represented by the predicted colors

We define two weight matrices

and

Our network optimizes over these two matrices with the loss L defined as the Euclidean distance between the prediction

and the ground-truth colors CΛt\s :







where the softmax function is applied column-wise. Intuitively, any visible target vertex, without color, will have it assigned to the weighted mean (softmax function) of all the available colored vertex set, with weights encoded in matrices W1 and W2. We interpolate the predicted vertex colors

from the learned model over the corresponding mesh faces Ft\s), project the faces, and obtain the missing regions in Is→t.

[0035] Generating Training Samples. The training data consists of inputs, each being a subset of colored vertices from a mesh, and outputs that represent different subsets of vertices from the same mesh. In practice, given any monocular image, once we fit the 3d model, we can generate any possible input-output split over the visible vertices. However, our inputs tend to be structured, consisting of subsets of vertices seen in the source and target mesh as well as their difference set. To ensure a similar distribution, we take the inputs and outputs to be sets of visible vertices in the intersection of the source and target mesh (we assume intersection is non-trivial), and choose outputs in their difference sets, respectively. Two training examples can thus be generated, symmetrically, for every pair of images of people, with different appearance and in a different pose.

[0036] The drawback of this procedure is that, at training time, the network has to predict colors from a smaller intersection set of colored vertices (i.e. ΛsΛt), whereas at test time, it can use the fully colored set of vertices from the source mesh Cs.

Clothing Layout Warping



[0037] We use the model of K. Gong, X. Liang, X. Shen, and L. Lin: "Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing", in CVPR, July 2017 to estimate the clothing layout for target and source images, Lt and Ls, defined over a set of 20 clothing labels. Given the clothing layout source Ls, we want to transform it into the pose of the target image, Ls⇒t. We start by collecting clothing label information for each visible vertex in the source mesh Ss. We propagate the labeling on the entire mesh by using a geodesic nearest-neighbor approach. For labels not on the source mesh, we collect the nearest-neighbour source vertex projections, vote on a displacement and translate the labeling accordingly. Thus, we obtain a rough estimate of Ls⇒t, which will denote by Ls→t. We gather additional image data from the web, consisting of source-target image pairs depicting the same person wearing the same clothes, but in different poses. On this dataset of ∼ 1; 500 training pairs, we train an image to image translation network which outputs Ls⇒t given Ls⇒t and the disparity map Dt.

[0038] Fig. 6 shows clothing layout warping. From left to right: source image clothing layout Ls, target image clothing layout Lt, input of the clothing layout synthesis network Ls→t, final output of the warping network Ls⇒t.

Human Appearance Synthesis



[0039] The previously described prediction Is→t captures the appearance transfer only as covered by our 3d body models. Hence, clothing layers (e.g. skirt, jacket) or hair which fall outside the coverage of the human body model are not transferred during the process. To achieve a higher perceptual quality for the generated image, we further refine our prediction using a Human Appearance Synthesis (HAS) network adapted based on ideas in Q. Chen and V. Koltun: "Photographic image synthesis with cascaded refinement networks" in ICCV, October 2017 (wherein the disclosure of said document are included herein by reference). This method performs multi-resolution refinement and was originally used in synthesizing photographic images conditioned on semantic segmentation layouts. Instead, we train the network to predict an image It given three types of inputs: a predicted semantic layout of clothing Lt or Ls ⇒ t, the disparity map Dt, and Is→t. The output of this HAS network, Is⇒t, represents our final, refined result.

[0040] Fig. 7 shows human appearance transfer with clothing layout warping. From left to right: source image, target image, clothing warped Ls⇒t and RGB data, Is⇒t, generated using the human appearance synthesis module.

EXPERIMENTS



[0041] For all of our experiments we use the Chictopia10k dataset (see X. Liang, C. Xu, X. Shen, J. Yang, S. Liu, J. Tang, L. Lin, and S. Yan: "Human parsing with contextualized convolutional neural network"; in ICCV, 2015; wherein the disclosure of said document is incorporated herein by reference). The images in this dataset depict different people, under both full and partial viewing, captured frontally. The high variability in color, clothing, illumination and pose makes this dataset suitable for our task. There are 17,706 images available together with additional ground truth clothing segmentations. We do not use the clothing labels provided, but only the figure-ground segmentation such that we can generate training images cropped on the human silhouette.

[0042] We split the data in two subsets: 15,404 images for training and 2,302 images for testing. We additionally prune some of the images based on the quality of body shape fitting, This is done by applying a soft threshold on the intersection over union (IoU) between the projection of the fitted body model and the foreground mask of the person. For each image in the dataset, we randomly select two other images from its corresponding subset (i.e. train or test) to construct image pairs. In the end, we use 28,808 training pairs and 4,080 testing pairs.

[0043] Appearance Transfer. Results of our Body Color Completion module are shown in Fig. 9. Sample results of our Human Appearance Synthesis module are also given in Fig. 10. Although we show transfer results in one direction, our method is symmetrical, so we obtain results of similar quality both ways, as shown in Fig. 1.

[0044] Fig. 9 shows sample results for the body color completion module. From left to right: source image, target image, appearance data copied using visible mesh triangles on both meshes (source and target), appearance data completed using vertex color completion.

[0045] Fig. 10 shows sample results for the human appearance transfer pipeline. From left to right: source image with corresponding fitted 3d body model, target image with its clothing layout and fitted 3d body model, RGB data generated using the Body Color Completion module (i.e. Is→t), RGB data generated using the Human Appearance Synthesis module (i.e. Is ⇒ t).

[0046] Impact of Components and Failure Modes. Our Human Appearance Synthesis network receives as inputs Is→t, the depth map and the clothing segmentation of the target image. To evaluate the contribution of each of these inputs in the visual quality of the output, we train two additional Human Appearance Synthesis networks under similar conditions, but with different input data: one without the depth map, and the other without both the depth map and the clothing segmentation. In Fig. 11, we provide visual results for all three networks. We observe that the best quality is obtained when using the complete network.

[0047] Fig. 11 shows impact of the different model components in the effectiveness of human appearance transfer. In the third column we show our complete method, in the fourth column we do not use the depth map and in the fifth column we neither use the depth map nor the clothing segmentation. For the first example (first row), notice the difference in quality for the girl's skirt (i.e. for the complete network the skirt is fully defined, for the network without depth the skirt becomes blurry and for the network without both depth and clothing, the skirt is nearly in-existent and body detail is blurred). For the second example (second row) notice the improved quality of the generated hair.

[0048] Errors occur in our pipeline when the clothing segmentation fails or the 3d body shape fitting does not yield good alignment with the person in the image. Examples are shown in Fig. 12.

[0049] Fig. 12 shows several failure modes of our method. From left to right: source image Is with corresponding mesh, target image It with corresponding mesh and the generated Is⇒t. In the first example the body shape fitting in Is misses the model's left hand, causing the final output to contain 3 hands, one resulting from barycentric transfer. In the second example, the model has problems dealing with the coat of the person from the source image. In the third example, the issue is caused by the small number of common vertices between the two body shape models, producing appearance mixing and considerable transfer ambiguity.

Identity Preserving Appearance Transfer



[0050] We also implement a variation of our model in order to preserve the identity of the target subject during appearance transfer. To do so, we redefine the sets of vertices used in the Body Color Completion module. We start by identifying the skin and clothing image pixels of the target and source images by using the set of labels provided by the clothing segmentation model. We place the set of labels defined by hair, face, arms and legs in the skin category and the remaining ones in the clothing category. Then we assign a category (i.e. skin/cloth) for each vertex in the body models by inspecting the clothing labeling under their projections in the image. We fix the colors for the pixels/vertices categorized as skin in the target, and perform barycentric transfer only for the intersection of source and target vertices categorized as clothing. The colors for the remaining vertices are predicted as before by our Vertex Color Completion network. Sample results are shown in Figs. 8a and 8b.

[0051] Figs. 8a and 8b show sample appearance transfer results from Is to It, where we preserved the identity of the person from It. From left to right: source image, Is, target image, It, generated image with person from It using the clothing of person from Is.

Perceptual User Study and Automated Tests



[0052] We perform a perceptual study by asking 20 human subjects to evaluate our results. We present each one with 100 results in the form of source image (Is), target image (It) and our automatically generated appearance transfer image Is⇒t. We ask subjects to evaluate the appearance transfer quality of Is⇒t, by assigning it one of the following scores: very poor (1), poor (2), good (3), very good (4), perfect (5). Finally, we quantified their scores in the form of a normalized histogram, shown in Fig. 13 (bottom). The mean score is 2.9, with standard deviation 0.96, suggesting that our transfer is reasonable on average.

[0053] Fig. 13 shows (top) distributions of person detection scores given by running Faster R-CNN (see S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: "Towards real-time object detection with region proposal networks", in NIPS, pages 91-99, 2015; wherein the disclosure of this document is incorporated herein by reference) over two sets of real and automatically generated images, respectively. (Bottom) We conducted a quality survey for a set of 100 automatically generated images by our method (randomly selected), where people were asked to assign a score from 1 (very poor quality) to 5 (perfect quality).

[0054] We compare our method against recent work (see L. Ma, X. Jia, Q. Sun, B. Schiele, T. Tuytelaars, and L. Van Gool: "Pose guided person image generation", in NIPS, 2017; A. Siarohin, E. Sangineto, S. Lathuiliere, and N. Sebe: "Deformable gans for pose-based human image generation", in CVPR, 2018; disclosures of both documents are incorporated herein by reference) independently addressing the problem of pose conditioned human image generation. Such methods rely solely on information in the image without explicitly inferring 3d body pose. Set aside significant methodological differences, we are additionally able to perform identity-preserving transfer (see Fig. 8). Our results are more visually pleasing and superior in terms of Inception Scores (see T. Salimans, I. Goodfellow,W. Zaremba, V. Cheung, A. Radford, X. Chen, and X. Chen: "Improved techniques for training gans", in NIPS, 2016; wherein the disclosure of said document is incorporated herein by reference), which are 3.09 (see L. Ma, X. Jia, Q. Sun, B. Schiele, T. Tuytelaars, and L. Van Gool: "Pose guided person image generation", in NIPS, 2017; wherein the disclosure of said document is incorporated herein by reference), 3.35 (see A. Siarohin, E. Sangineto, S. Lathuiliere, and N. Sebe: "Deformable gans for pose-based human image generation", in CVPR, 2018; wherein the disclosure of said document is incorporated herein by reference) and 4.13 (Ours).

[0055] In order to understand possible difference in terms of image statistics, we also perform an automated test using a state-of-the-art human detector, Faster R-CNN (see S. Ren, K. He, R. Girshick, and J. Sun: "Faster r-cnn: Towards real-time object detection with region proposal networks", in NIPS, pages 91-99, 2015; wherein the disclosure of said document is incorporated herein by reference). We compute the human detection scores on two sets containing 2,000 generated and real images, respectively. In Fig. 13 (top) we observe that the two distributions of detection scores are similar, with a dominant mode around value 0.99.

CONCLUSIONS



[0056] Modeling and synthesizing human appearance is difficult due to variability in human body proportions, shape, clothing and poses. However, models that can realistically synthesize complete images of humans under a degree of control (conditioning) on an input image appearance or pose, could be valuable for entertainment, photo-editing, or affordable online shopping of clothing. In this context, we define a new problem entitled human appearance transfer where given two images, source and target, of different people with different poses and clothing, we learn to transfer the appearance of the source person on the body layout of the target person. Our solution relies on state-of-the-art 3d human pose and shape estimation based on deep multitask neural networks and parametric human shape modeling, combined with deep photographic synthesis networks controlled by appropriate 2d and 3d inputs. Our image results, backed-up by a perceptual user study, Inception scores, and the response of a state-of-the-art human person detector indicate that the proposed model can automatically generate images of humans of good perceptual quality, and with similar statistics as real human images. We also show how the model can be modified to realistically 'dress' a person shown in one image with clothing captured from a person in another image.

MONOCULAR 3D POSE AND SHAPE ESTIMATION OF MULTIPLE PEOPLE IN NATURAL SCENES



[0057] In the following, monocular 3d pose and shape estimation of multiple people in natural scenes will be considered in more detail and supplement the aforesaid.

[0058] Human sensing has greatly benefited from recent advances in deep learning, parametric human modeling, and large scale 2d and 3d datasets. However, existing 3d models make strong assumptions about the scene, considering either a single person per image, full views of the person, a simple background or many cameras. In this paper, we leverage state-of-the-art deep multi-task neural networks and parametric human and scene modeling, towards a fully automatic monocular visual sensing system for multiple interacting people, which (i) infers the 2d and 3d pose and shape of multiple people from a single image, relying on detailed semantic representations at both model and image level, to guide a combined optimization with feedforward and feedback components, (ii) automatically integrates scene constraints including ground plane support and simultaneous volume occupancy by multiple people, and (iii) extends the single image model to video by optimally solving the temporal person assignment problem and imposing coherent temporal pose and motion reconstructions while preserving image alignment fidelity. We perform experiments on both single and multi-person datasets, and systematically evaluate each component of the model, showing improved performance and extensive multiple human sensing capability. We also apply our method to images with multiple people, severe occlusions and diverse backgrounds captured in challenging natural scenes, and obtain results of good perceptual quality.

[0059] Accurately detecting and reconstructing multiple people, possibly involved in interactions with each other and with the scene, based on images and video data, has extensive applications in areas as diverse as human-computer interaction, human behavioral modeling, assisted therapy, monitoring sports performances, protection and security, special effects, modeling and indexing archival footage, or self-driving cars.

[0060] To support the level of modeling accuracy required by such applications, we ultimately need highly-detailed models able not just to detect people and their body joints in images, but also the spatial extent of body parts, as well as the threedimensional pose, shape and motion for each person in the scene. For complex scenes, such demands would likely require a virtuous cycle between 2d and 3d reasoning, with feedback. One should further consider integrating anthropometry constraints, avoiding geometric collisions between the estimated models of multiple people, and reasoning about ground planes implicit in many scenes, as people rarely float, unsupported in space - and if so, usually not for long. Reconstructions must also be temporally fluid and humanly plausible. Most importantly, constraints need to be enforced in the context of an image observation process which - even with many cameras pointed at the scene - remains incomplete and uncertain, especially in scenarios where multiple people interact. While the integration of such constraints appears challenging, their use provides the opportunity to considerably restrict the degrees of freedom of any natural human parameterization towards plausible solutions.

[0061] Herein, we address the monocular inference problem for multiple interacting people, by providing a model for 2d and 3d pose and shape reconstruction over time. Our contributions include (i) a semantic feedforward-feedback module that combines 2d human joint detection, semantic segmentation, and 3d pose prediction of people, with pose and shape refinement based on a novel semantic cost that aligns the model body parts with their corresponding semantic images regions, producing solutions that explain the complete person layout while taking into account its estimation uncertainty, (ii) incorporation of scene consistency measures including automatic estimation and integration of ground plane constraints, as well as adaptively avoiding simultaneous volume occupancy by several people, and (iii) resolution of the temporal person assignment problem based on body shape, appearance and motion cues within a Hungarian matching method, then solving a joint multiple-person smoothing problem under both 2d projection and 3d pose temporal fluidity constraints. Our quantitative results on datasets like Panoptic (see H. Joo, H. Liu, L. Tan, L. Gui, B. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Y. Sheikh: "Panoptic studio: A massively multiview system for social motion capture", in ICCV, 2015; wherein the disclosure of said document is incorporated herein by reference) and Human3.6M (see C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu: "Human3.6M: Large scale datasets and predictive methods for 3d human sensing in natural environments", PAMI, 2014; wherein the disclosure of said document is incorporated by reference) validate the importance of the ingredients in the proposed design. Qualitative results in complex monocular images and video show that the model is able to reconstruct multiple interacting people in challenging scenes in a perceptually plausible way. The model also supports the realistic synthesis of human clothing and appearance (human appearance transfer) as shown in our companion paper (see M. Zanfir, A. Popa, and C. Sminchisescu: "Human appearance transfer", in CVPR, 2018; wherein the disclosure of said document is incorporated herein by reference).

[0062] Fig. 14 shows an automatic 3d reconstruction of the pose and shape of multiple people from a monocular image, as estimated by our model integrating person and scene constraints. We leverage feedforward and semantic feedback calculations for each person, with joint reasoning on ground plane and volume occupancy over all the people in the scene.

[0063] Our work relates to recently developed deep architectures for 2d human pose estimation (see Z. Cao, T. Simon, S. Wei, and Y. Sheikh: "Realtime multiperson 2d pose estimation using part affinity fields", in CVPR, 2017; E. Insafutdinov, L. Pishchulin, B. Andres, M. Andriluka, and B. Schiele: "DeeperCut: A deeper, stronger, and faster multiperson pose estimation model", in ECCV, 2016; G. Papandreou, T. Zhu, N. Kanazawa, A. Toshev, J. Tompson, C. Bregler, and K. Murphy: "Towards accurate multi-person pose estimation in the wild", in CVPR, 2017; J. J. Tompson, A. Jain, Y. LeCun, and C. Bregler: "Joint training of a convolutional network and a graphical model for human pose estimation", in NIPS, 2014; A. Toshev and C. Szegedy. Deeppose: "Human pose estimation via deep neural networks", in CVPR, 2014; wherein the disclosures of said documents are incorporated herein by reference), 3d human pose estimation based on fitting volumetric models (see F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016; M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black: "SMPL: A skinned multi-person linear model", SIGGRAPH, 34(6):248:1-16, 2015; wherein the disclosures of said references are included herein by reference), feedforward deep models for 3d prediction (see J. Martinez, R. Hossain, J. Romero, and J. J. Little: "A simple yet effective baseline for 3d human pose estimation", in ICCV, 2017; G. Pavlakos, X. Zhou, K. G. Derpanis, and K. Daniilidis: "Coarse-to-fine volumetric prediction for single-image 3d human pose", in CVPR, 2017; X. Zhou, M. Zhu, K. Derpanis, and K. Daniilidis: "Sparseness meets deepness: 3d human pose estimation from monocular video", in CVPR, 2016; wherein the disclosures of said references are included herein by reference), as well as integrated deep models for 2d and 3d reasoning (see A. Popa, M. Zanfir, and C. Sminchisescu: "Deep multitask architecture for integrated 2d and 3d human sensing", in CVPR, 2017; G. Rogez and C. Schmid: "Mocap-guided data augmentation for 3d pose estimation in the wild", in NIPS, 2016; B. Tekin, P. Marquez Neila, M. Salzmann, and P. Fua: "Learning to fuse 2d and 3d image cues for monocular body pose estimation", in ICCV, 2017; D. Mehta, S. Sridhar, O. Sotnychenko, H. Rhodin, M. Shafiei, H.-P. Seidel, W. Xu, D. Casas, and C. Theobalt: "Vnect: Realtime 3d human pose estimation with a single rgb camera", ACM Transactions on Graphics (TOG), 36(4):44, 2017; wherein the disclosures of said documents are incorporated herein by reference). Accurate shape and motion-capture systems, based on multiple cameras or simplified backgrounds, have also been proposed with impressive reconstruction results (see A. Boukhayma, J.-S. Franco, and E. Boyer: "Surface motion capture transfer with Gaussian process regression", in CVPR, 2017; D. A. Forsyth, O. Arikan, L. Ikemoto, J. O'Brien, and D. Ramanan: "Computational Studies of Human Motion: Tracking and Motion Synthesis", NOW Publishers Inc, 2006; V. Leroy, J.-S. Franco, and E. Boyer: "Multi-view dynamic shape refinement using local temporal integration", in ICCV, 2017; H. Rhodin, N. Robertini, D. Casas, C. Richardt, H.-P. Seidel, and C. Theobalt: "General automatic human shape and motion capture using volumetric contour cues", in ECCV, 2016; wherein the disclosures of said documents are incorporated herein by reference). Systems designed for the 3d reconstruction of multiple people are relatively rare and existing ones are based on multiple cameras (see V. Belagiannis, S. Amin, M. Andriluka, B. Schiele, N. Navab, and S. Ilic: "3d pictorial structures for multiple human pose estimation", in CVPR, 2014; A. Elhayek, E. de Aguiar, A. Jain, J. Thompson, L. Pishchulin, M. Andriluka, C. Bregler, B. Schiele, and C. Theobalt: "Marconi-convnet-based marker-less motion capture in outdoor and indoor scenes", IEEE transactions on pattern analysis and machine intelligence, 39(3):501-514, 2017; A. Elhayek, E. Aguiar, A. Jain, J. Tompson, L. Pishchulin, M. Andriluka, C. Bregler, B. Schiele, and C. Theobalt: "Efficient ConvNet-based marker-less motion capture in general scenes with a low number of cameras", in CVPR, 2015; H. Joo, H. Liu, L. Tan, L. Gui, B. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Y. Sheikh: "Panoptic studio: A massively multiview system for social motion capture", in ICCV, 2015; Y. Liu, J. Gall, C. Stoll, Q. Dai, H.-P. Seidel, and C. Theobalt: "Markerless motion capture of multiple characters using Multiview image segmentation", IEEE transactions on pattern analysis and machine intelligence, 35(11):2720-2735, 2013; wherein the disclosures of said documents are incorporated herein by reference). In A. Elhayek, E. de Aguiar, A. Jain, J. Thompson, L. Pishchulin, M. Andriluka, C. Bregler, B. Schiele, and C. Theobalt: "Marconi-convnet-based marker-less motion capture in outdoor and indoor scenes", IEEE transactions on pattern analysis and machine intelligence, 39(3):501-514, 2017 (wherein the disclosure of said document is incorporated herein by reference), the method uses an arguably low number of cameras (3-4) to reconstruct several people, with promising results, but the level of interaction is somewhat limited. The work of H. Joo, H. Liu, L. Tan, L. Gui, B. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Y. Sheikh: "Panoptic studio: A massively multiview system for social motion capture", in ICCV, 2015 (wherein the disclosure of said document is incorporated herein by reference) proposes a multi-person tracking system (which we also use for our 'ground-truth' monocular evaluation), although the system relies on a massive number of RGB and RGB-D cameras for inference, and the capture dome offers inherently limited background variability. Our single person initialization relies on the Deep Multitask Human Sensing Network (DMHS) (see A. Popa, M. Zanfir, and C. Sminchisescu: "Deep multitask architecture for integrated 2d and 3d human sensing", in CVPR, 2017; wherein the disclosure of said reference is incorporated herein by reference) for initial 2d and 3d pose inference (body joints, semantic segmentation, pose prediction), which is then refined based on our own implementation of the human body model SMPL (see M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black: "SMPL: A skinned multi-person linear model", SIGGRAPH, 34(6):248:1-16, 2015; wherein the disclosure of said document is incorporated herein by reference), augmented with learned semantic vertex labeling information, and using a new semantic loss function, which represents one of our contributions. Systems based on discriminative-generative (feedforward-feedback) components for 3d human pose estimation date, in principle, back at least to D. Ramanan and C. Sminchisescu: "Training deformable models for localization", in CVPR, 2006; L. Sigal, A. Balan, and M. J. Black: "Combined discriminative and generative articulated pose and non-rigid shape estimation", in NIPS, 2007; C. Sminchisescu, A. Kanaujia, and D. Metaxas: "Learning joint top-down and bottom-up processes for 3d visual inference", in CVPR, 2006 (wherein the disclosures of said documents are incorporated herein by reference) but our approach leverages considerably different image representations, body models, cost functions and optimization techniques. Our automatic ground plane and adaptive people volume occupancy exclusion constraints, as well as our multiple people assignment and smoothing costs are integrated in a novel and coherent way, although monocular single person costs based on simpler model formulations and/or multiple hypotheses tracking techniques exist in the literature (see F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016; R. M. Neal, M. J. Beal, and S. T. Roweis: "Inferring state sequences for non-linear systems with embedded hidden Markov models", in NIPS, 2004; V. Ramakrishna, T. Kanade, and Y. Sheikh: "Reconstructing 3d human pose from 2d image landmarks", ECCV, 2012; E. Simo-Serra, A. Quattoni, C. Torras, and F. Moreno-Noguer: "Ajoint model for 2d and 3d pose estimation from a single image", in CVPR, 2013; C. Sminchisescu and A. Jepson: "Variational mixture smoothing for non-linear dynamical systems", in CVPR, 2004; H. Yasin, U. Iqbal, B. Kruger, A. Weber, and J. Gall: "A dualsource approach for 3d pose estimation from a single image", in CVPR, 2016; wherein the disclosures of said documents are incorporated herein by reference).

Multiple Persons in the Scene Model



[0064] Problem formulation. Without loss of generality, we consider Np uniquely detected persons in a video with Nf frames. Our objective is to infer the best pose state variables

shape parameters

and individual person translations

with PNp and fNf. We start by first writing a per-frame, person-centric objective function



where the cost LS takes into account the visual evidence computed in every frame in the form of semantic body part labeling, LC penalizes simultaneous (3d) volume occupancy between different people in the scene, and LG incorporates the constraint that some of the people in the scene may have a common supporting plane. The term

is a Gaussian mixture prior similar to F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016 (wherein the disclosure of said document is incorporated herein by reference). The image cost for multiple people under all constraints can be written as



[0065] If a monocular video is available, the static cost Lf is augmented with a trajectory model applicable to each person once the temporal assignment throughout the entire video has been resolved. The complete video loss writes

where LT can incorporate prior knowledge on human motion, ranging from smoothness, assumptions of constant velocity or acceleration, or more sophisticated models learned from human motion capture data. In the next sections, we describe each cost function in detail. Whenever unambiguous, we drop the f and p super-scripts.

[0066] In order to infer the pose and 3d position of multiple people we rely on a parametric human representation, SMPL (see M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black: "SMPL: A skinned multi-person linear model", SIGGRAPH, 34(6):248:1-16, 2015; wherein the disclosure of said document is incorporated herein by reference), with a state-of-the-art deep multitask neural network for human sensing, DMHS (see A. Popa, M. Zanfir, and C. Sminchisescu: "Deep multitask architecture for integrated 2d and 3d human sensing", in CVPR, 2017; wherein the disclosure of said document is incorporated herein by reference). In practice, we cannot assume a constant number of people throughout a video and we first infer the parameters B,Θ,T independently for each frame by minimizing the sum of the first two cost functions: LS and LC. Then, we temporally track the persons obtained in each frame by means of optimally solving an assignment problem, then re-optimize the objective, by adding the temporal and ground plane constraints, LT and LG. For those people detected in only some of the frames, optimization will proceed accordingly over the corresponding subset. An overview of the method is shown in Fig. 15.

[0067] Fig. 15 shows a processing pipeline of our monocular model for the estimation of 3d pose and body shape of multiple people. The system combines a single person model that incorporates feedforward initialization and semantic feedback, with additional constraints such as ground plane estimation, mutual volume exclusion, and joint inference for all people in the scene. For monocular video, the 3d temporal assignment of people is resolved using a Hungarian method, and trajectory optimization is performed jointly over all people and timesteps, under all constraints, including image consistency, for optimal results.

Single Person Feedforward Feedback Model



[0068] SMPL (see M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black: "SMPL: A skinned multi-person linear model", SIGGRAPH, 34(6):248:1-16, 2015; wherein the disclosure of said document is incorporated herein by reference) is a differentiable parametric human model - represented by template vertices V0 - and controlled by pose vectors

and shape parameters

The pose of the model is defined by a standard skeletal rig that has the main body joints. For each body part, the vectors controlling the pose are provided in axis-angle representations of the relative rotations w.r.t. their parents in the kinematic tree. The axis angle for every joint is transformed to a rotation matrix using the Rodrigues transformation. The shape parameters β impact limb size, height and weight and represent coefficients of a low dimensional shape space learned from registered meshes. SMPL provides matrix functions dependent on θ and β, namely ,

which gives the transformed vertex positions for the whole mesh, and

which outputs the joint positions for the associated kinematic tree. The total number of vertices in the SMPL model is nV = 6890 and the total number of joints in the kinematic tree is nJ = 24. For simplicity of explanation, let v denote V(θ,β|Vo) and let x be J(θ,β|Vo). We refer to the translation of the model in camera space as



[0069] DMHS (see A. Popa, M. Zanfir, and C. Sminchisescu: "Deep multitask architecture for integrated 2d and 3d human sensing", in CVPR, 2017; wherein the disclosure of said document is incorporated herein by reference) is a state-of-the-art feedforward multi-task deep neural network for human sensing that provides, for a given image

the following estimates: the 2d and 3d joints of a single person as well as the semantic body parts at pixel level. We denote these 3 outputs by the matrices



and

respectively. We denote by mJ = 17 the number of joints in the representation considered by the network and Ns = 25 the number of semantic body parts. The method has been shown to perform well for both indoor images as well as outdoor. The challenges of integrating DMHS and SMPL stem from accurately fitting (transferring) the parametric SMPL model to the 3d joint positions predicted by DMHS, as well as designing semantic-based cost functions that allow to efficiently couple the model to the observations - perform 3d fitting in order to best explain the human layout in the image. In order to semantically assign model mesh components to corresponding image regions, one needs a consistent 'coloring' of their vertices according to the NS human body part labels available e.g. in Human3.6M (see C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu: "Human3.6M: Large scale datasets and predictive methods for 3d human sensing in natural environments", PAMI, 2014; wherein the disclosure of said document is included herein by reference). This can be achieved robustly, during a training process. We project and fit the SMPL model in multiple (4) views and for different ground truth poses from Human3.6M (we chose 100 different poses). Then each model vertex was associated the median image body part label, available in Human3.6M, transferred from images to the corresponding vertex projections. See Fig. 17 for coloring examples.

[0070] Fig. 17 shows importance of semantic feedback in capturing the correct 3d pose and body shape. From left to right: input image, semantic body labeling produced by DMHS, inferred body shape and pose without the semantic term (ΦJ only) and the semantically fitted model ΦS.

[0071] Single Person Feedforward Feedback Model: Feedforward Prediction, Pose & Shape Transfer

[0072] We detail the transfer procedure for a single person and perform the same steps for all people in each frame of a video. To transfer the feedforward prediction of the configuration of joints y3D obtained from DMHS to the SMPL model, we have to define a cost function Φ3d(θ,β), and infer optimal θ and β parameters. One such cost function is the Euclidean distance between joints shared in both representations (i.e. i,jCJ, where 1 ≤ i ≤ mJ and CJ is the set of compatible joint indices)

where h indicates the index of the pelvis and x(j) - x(h) represents the centered 3d pose configuration with respect to the pelvis joint. Unless otherwise stated, we use ∥·∥ for the l2 norm, ∥·∥2.

[0073] However, based on (4) the DMHS to SMPL transfer is unsatisfactory. This is because 1) the prediction made by DMHS is not necessarily a valid human shape, and 2) a configuration in the parameter space of β or even in the space of θ does not necessarily represent an anatomically correct human pose. In F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016 (wherein the disclosure of said document in included herein by reference), multiple regularizers were added: a norm penalty on β and a prior distribution on θ. However, these risk excessive bias.

[0074] We propose an alternative transfer equation, focusing on qualitatively modeling the pose predicted by DMHS so to preserve the 3d orientation of limbs. Our function Φcos penalizes the cosine distance between limbs - or selected pairs of joints - that are shared in both representations (property denoted by (i,j), (a, b) ∈ CL, wherein 1 ≤ i,jmJ and 1 ≤ k, lnJ. Given aij = y3D(i) - y3D(i) and bkl = x(k) - x(l), the cost is



[0075] While in practice the minimization of Φcos converges quickly to a perfect solution (often close to 0) and the resulting pose is perceptually similar to DMHS, the implicit shape information provided by DMHS is lost. In situations where the original 3d joint prediction confidence is high (e.g. training and testing distributions are expected to be similar, as in Human3.6M), one can further optimize over β, starting from solutions of (5)



Results of the proposed transfer variants are shown in fig. 16.

[0076] Fig. 16 shows a 3d pose transfer from DMHS to SMPL: (a) input image; (b) 3d joints with links, as estimated by DMHS; (c) transfer after applying (4) directly minimizing Euclidean distances between common 3d joints in both representations. Notice unnatural body shape and weak perceptual resemblance with the DMHS output. (d) is also obtained using (4) but with extra regularization on pose angles - offering plausible configurations but weak fits. (e) transfer results obtained using our proposed cost (5) which preserves limb orientation, and (f) inferred configurations after our semantic optimization, initialized by (e).

Single Person Feedforward Feedback Model: Semantic 3d Pose and Shape Feedback



[0077] After transferring the pose from DMHS to SMPL we obtain an initial set of parameters θ0 and β0 and one can refine the initial DMHS estimate. One way to fit the 3d pose and shape model starting from an initialization (see F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black: "Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image", in ECCV, 2016; C. Sminchisescu and B. Triggs: "Estimating Articulated Human Motion with Covariance Scaled Sampling", IJRR, 22(6):371-393, 2003; wherein the disclosures of said document are included herein by reference), is to minimize the projection error between the model joints and the corresponding detected image joint locations, y2d. We denote by

the image projection function, with fixed camera intrinsincs. One possible loss is the Euclidean distance, computed over sparse joint sets weighted by their detection confidence w (some may not be visible at all)



[0078] The problem of minimizing θ and β for monocular error functions, defined over distances between sparse sets of joints, is its ambiguity, as the system is clearly underdetermined, especially for depth related state space directions that couple along camera's ray of sight (see C. Sminchisescu and B. Triggs: "Kinematic jump processes for monocular 3d human tracking", in CVPR, 2003; wherein the disclosure of said document is incorporated herein by reference). We propose a new error function based on projecting the mesh v in the image I and measuring the dense, pixel-wise semantic error between the semantic segmentation transferred by the model projection and a given DMHS semantic body part segmentation prediction ys.

[0079] We are given Ns semantic classes that describe body parts with ys storing semantic confidence maps. We construct a function fS(p = (x,y)T) = argmaxk yS(p,k) with 1 ≤ xW, 1 ≤ yH integers, that returns the body part label 1 ≤ k ≤ NS of pixel location p in the image I. Let vk be vertices pertaining to the body part indexed in k and

their image projection.

[0080] We design a cost ΦS(Θ,B,T), where each point p from the semantic body part segmentation maps finds its nearest neighbour in pk, and drags it in place. Appropriately using pixel label confidences (x, y) for a given class k as yS is important for robust estimates in a cost that writes



[0081] In practice, our semantic cost is further weighted by a normalization factor 1/Z, with

ensuring φS remains stable to scale transformations impacting the area of the semantic map (closer or further away, with larger or smaller number of pixels, respectively). Another desirable property of the semantic loss is that when confidences are small, ΦS will have a lower weight in the total loss, emphasizing other qualitatively different terms in the cost. The total semantic loss can then be written


Simultaneous Volume Occupancy Exclusion



[0082] To ensure that estimated models of people in a scene are not inconsistent, by occupying the same 3d space volume simultaneously, we need additional processing. We design adaptive representations to first compute enclosing parallelepipeds for each person according to its current model estimates, rapidly test for intersections (far-range check), and only integrate detailed, close range collision avoidance into the loss when the far-range response is negative. For close-range volume occupancy exclusion, we use specialized terms obtained as follows: for each person model, we fit tapered superquadrics to each limb, and represent the limb by a series of Nb fitted spheres inside the superquadric, with centers c and radius r. For any two persons, p and p', we define the loss LC(p, p') based on distances between all spheres belonging, respectively, to the first and second person





[0083] The loss for Np persons in a frame is defined as the sum over all pair-wise close-range losses LC(p, p') among people with negative far-range tests. People with positive far-range tests do not contribute to the volume occupancy loss. Notice how this cost potentially couples parameters from all people and requires access to their estimates. See Fig. 18 for visual illustrations.

[0084] Fig. 18 shows adaptive volume occupancy avoidance: (a) input image where people are far apart; (b) visual representation for far-range collision check; (c) image where people are in contact; (d) inferred body shapes without and (e) with collision constraint, which ensures correct contact without simultaneous volume occupancy.

Ground Plane Estimation and Constraint



[0085] We include a prior that the scene has a ground-plane on which, on average, the subjects stand and perform actions. To build a correct hypothesis for the location and orientation of the plane, we design a cost that models interactions between the plane and all human subjects, but leaves room for outliers, including people who, temporarily or permanently, are not in contact with the ground. Specifically, we select the 3d ankle positions of all persons in all the frames of a video, be these xi, and fit a plane to their locations.

[0086] We assume that a point z is on a plane with a surface normal n if the following equation is satisfied (z - p)T n = 0, where p is any fixed point on the plane. Given that some of the ankles might be occluded, we use a confidence term to describe the impact they have on the fitting process. We use the confidence wi from the DMHS 2d joint detector, with a two-folded purpose to 1) select the initial point p belonging to the plane as the weighted median of all ankle locations of the detected persons, and 2) weight measurements used in the robust L1 norm estimate of the plane hypothesis. Our plane estimation objective is



[0087] The estimates (p, n*) are then used in the ground-plane constraint term LG to penalize configurations with 3d ankle joints estimates away from the plane



[0088] Where the subscripts 1 and r identify the left and the right ankles for a person p at time f. The weighting of the terms is performed adaptively based on confidences w1, wr of the associated ankle joints. If these are not visible, or are visible within some distance of the ground and not confident, constraints are applied. If the joints are visible and confident, or far from the ground, constraints are not applied.

Assignment and Trajectory Optimization



[0089] Independently performing 3d human body pose and shape optimization in a monocular video can lead to large translation variations along depth directions and movements that lack natural smoothness. For this reason, we propose a temporal constraint that ensures for each of the inferred models that estimates in adjacent frames are smooth. To achieve it, we first need to resolve the assignment problem over time (identify or track the same individual throughout the video), then perform temporal smoothing for each individual track.

[0090] To solve the person assignment problem, we use the Hungarian algorithm to optimally build tracks based on an interframe inter-person cost combining the appearance consistency (measured as distances between vectors containing the median colors of the different body parts, computed over the model vertices), the body shape similarity, and the distance between the appropriately translated 3d joints inferred for each person, at frame level.

[0091] Once the assignment has been resolved between every pair of estimated person models in every successive set of frames, and tracks are built, several motion priors can be used, ranging from a constant velocity model, to more sophisticated auto-regressive processes or deep recursive predictors learned from training data (see J. Martinez, M. J. Black, and J. Romero: "On human motion prediction using recurrent neural networks", CoRR, abs/1705.02445, 2017; K. Fragkiadaki, S. Levine, P. Felsen, and J. Malik: "Recurrent network models for human dynamics", in ICCV, pages 4346-4354, 2015; J. M. Wang, D. J. Fleet, and A. Hertzmann: "Gaussian process dynamical models", in NIPS, 2006; wherein the disclosures of said documents are incorporated herein by reference. The integration of such motion representations in our framework is straightforward as long as they remain differentiable. Here we experiment with constant velocity priors on pose angles, Θ as well as translation variables, T. Our temporal loss function component at.

frame f ≥ 2 for a person (track) p is defined as



[0092] The shape parameters βp are set as the median of

f. Because smoothing axis-angle representations is difficult, the angle-related costs in (15) are represented using quaternions, which are easier to smooth. Gradients are propagated through the axis-angle to quaternion transformation during the optimization.

Experiments



[0093] We numerically test our inference method on two datasets, CMU Panoptic (see H. Joo, H. Liu, L. Tan, L. Gui, B. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Y. Sheikh: "Panoptic studio: A massively multiview system for social motion capture", in ICCV, 2015; wherein the disclosure of said document is incorporated herein by reference) and Human3.6M (see C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu: "Human3.6M: Large scale datasets and predictive methods for 3d human sensing in natural environments", PAMI, 2014; wherein the disclosure of said document is incorporated herein by reference), as well as qualitatively on challenging natural scenes (see Fig. 20). On Human3.6M we test different components of the model including semantic feedback, smoothing and the effect of multiview constraints. Panoptic in turn provides the real quantitative test-bed for the complete monocular system.

[0094] Fig. 20 shows an automatic 3d reconstruction of multiple people from monocular images of complex natural scenes. Left to right: input image, inferred model overlaid, and two different views of 3d reconstructions obtained by our model (including ground plane). Challenging poses, occlusions, different scales and close interactions are correctly resolved in the reconstruction.

[0095] Given a video with multiple people, we first detect the persons in each frame and obtain initial feedforward DMHS estimates for their 2d body joints, semantic segmentation and 3d pose. Similarly to E. Marinoiu, M. Zanfir, V. Olaru, and C. Sminchisescu: "3D Human Sensing, Action and Emotion Recognition in Robot Assisted Therapy of Children with Autism", in CVPR, 2018 (wherein the disclosure of said document is incorporated herein by reference), we extend DMHS to partially visible people, by fine-tuning both the semantic and the 3d pose estimation components of DMHS on a partial view version of Human80K (see C. Ionescu, J. Carreira, and C. Sminchisescu: "Iterated secondorder label sensitive pooling for 3d human pose estimation", in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1661-1668, 2014; wherein the disclosure of said document is incorporated herein by reference). For each person we perform the transfer proposed in (5) that aligns the limb directions of 3d estimates predicted by DMSH with the limb directions of SMPL. The transfer gives an initialization for pose and shape. The initial translation of each person is set to 3 meters in front of the camera.
Table 1: Automatic 3d human pose and translation estimation errors (in mm) on the Panoptic dataset (9,600 frames, 21,404 people). Notice the value of each component and the impact of the ground-plane constraint on correct translation estimation. Reference 23 behind the DMHS indicates A. Popa, M. Zanfir, and C. Sminchisescu: "Deep multitask architecture for integrated 2d and 3d human sensing", in CVPR, 2017, wherein the disclosure of said document is incorporated herein by reference.
 HagglingMafiaUltimatumPizzaMean
MethodPoseTranslationPoseTranslationPoseTranslationPoseTranslationPoseTranslation
DMHS [23] 217.9 - 187.3 - 193.6 - 221.3 - 203.1 -
2d Lose 135.1 282.3 174.5 502.2 143.6 357.6 177.8 419.3 157,7 390.3
Semantic Less 144.3 260.5 179.0 159.8 160.7 376.6 178.6 413.6 165.6 377.6
Smoothing 141.4 260.3 173.6 454.9 155.2 368.0 173.1 403.0 160.8 371.7
Smoothing Ground Plane 140.0 257.8 165.9 400.5 150.7 301.1 156.0 294.0 152.4 315.5
Table 2: Mean per joint 3d position error (in mm) on the Human3.6M dataset, evaluated on the test set of several very challenging actions. Notice the importance of various constraints in improving estimation error. Reference 23 behind the DMHS indicates A. Popa, M. Zanfir, and C. Sminchisescu: "Deep multitask architecture for integrated 2d and 3d human sensing", in CVPR, 2017, wherein the disclosure of said document is incorporated herein by reference.
MethodWalkingDogSittingSitting Down
DMHS [23] 78 119 106
Semantic Loss 75 109 101
Multi View 51 71 65
Smoothing 48 68 64


[0096] Human3.6M is a large-scale dataset that contains single person images recorded in a laboratory setup using a motion capture system. The dataset has been captured using 4 synchronized RGB cameras and contains videos of 11 actors performing different daily activities. We select 3 of the most difficult actions: sitting, sitting down and walking dog to test our single-person model. We use the official left-out test set from the selected actions, consisting of 160K examples. On this dataset we can only evaluate the pose inference under MPJPE error, but without the translation relative to the camera. We show results in table 2. We obtain an improvement over DMHS by using the proposed semantic 3d pose and shape feedback, cf. (10). On this dataset, we also experiment with multi-view inference and show a consistent improvement in 3d pose estimation. For multi-view inference, the loss function proposed in (10) is easily extended as a sum over measurements in all available cameras. Adding a temporal smoothness constraint further reduces the error. We also evaluated our method on all 15 actions from the official test set (911,744 configurations) and obtain an average error of 69 mm. Detailed results can be seen at http://vision.imar.ro/human3.6m/ranking.php (Testset H36M_NOS10).

[0097] CMU Panoptic Dataset. We selected data from 4 activities (Haggling, Mafia, Ultimatum and Pizza) which contain multiple people interacting with each other. For each activity we selected 2 sub-sequences, each lasting 20 seconds (i.e. 600 frames), from HD cameras indices 30 and 16 (for variability only, all testing is monocular). In total, we obtain 9,600 frames that contain 21,404 people. We do not validate/train any part of our method on this data.

[0098] Evaluation Procedure. We evaluate both the inferred pose, centered in its hip joint, under mean per joint position error (MPJPE), and the estimated translation for each person under standard Euclidean distance. We perform the evaluation for each frame in a sequence, and average the results across persons and frames. We match each ground-truth person in the scene with an estimation of our model. For every ground truth pose, we select the closest inferred model under the Euclidean distance, in camera space.

[0099] Ablation Studies. We systematically test the main components of the proposed monocular inference system and show the results detailed for each activity in table 1. Compared to DMHS, our complete method reduces the MPJPE error significantly, from 203.4 mm to 153.4 mm on average (-25%), while also computing the translation of each person in the scene. The translation error is, on average, 315.5 mm. The semantic projection term helps disambiguate the 3d position of persons and reduces the translation error compared to using only the 2d projection term. Temporally smoothing the pose estimates decreases the translation error further. Imposing the ground plane constraint makes the most significant contribution in this setup, decreasing the total translation error from 371 mm to 315 mm (-15%). Even though the total pose error also decreases when all constraints are imposed, on some sequences (e.g. Haggling) the error did not decrease when semantic terms are used. At a closer look, we noticed that the semantic maps and 3d initialization from DMHS were particularly noisy on those sequences of Haggling, camera index 30. Qualitative results in monocular images from the Panoptic dataset are shown in Fig. 19. Our method produces perceptually plausible 3d reconstructions with good image alignment in scenes with many people, some only partially visible, and captured under non-conventional viewing angles.

[0100] Fig. 19 shows an automatic monocular 3d reconstruction of multiple people in Panoptic videos. Left to right: input image, inferred model overlaid to assess fitting quality, two different views of the 3d reconstruction. Unusual viewing angles, pose variability, partial views and occlusions, make monocular reconstruction challenging. Quantitative results are given in table 1.

Conclusions



[0101] We have presented a monocular model for the integrated 2d and 3d pose and shape estimation of multiple people, under multiple scene constraints. The model relies on feedforward predictors for initialization and semantic fitting for feedback and precise refinement (shape adaption) to the observed person layout. It estimates and further integrates ground plane and volume occupancy constraints, as well as temporal priors for consistent, plausible estimates, within a single joint optimization problem over the combined representation of multiple people, in space and time. Our experimental evaluation, including ablation studies, is extensive, covers both single-person and multiple-person datasets and illustrates the importance of integrating multiple constraints.

[0102] Moreover, we qualitatively show that the method produces 3d reconstructions with tight image alignment and good perceptual quality, in both monocular images and video filmed in complex scenes, with multiple people, severe occlusion and challenging backgrounds. To our knowledge, such a large-scale fully automatic monocular system for multiple person sensing under scene constraints has been presented here for the first time.

[0103] It has to be noted that any of the embodiments and features of any of the embodiments, described herein, may be combined with each other, unless a combination is explicitly excluded.

[0104] Additionally, also other variations to the enclosed embodiments can be understood and effected by those skilled in the art and practicing the claimed invention, from a study of the drawings, the disclosure and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

[0105] In the following, there are provided further explanations of examples of the present invention headed "Human Appearance Transfer" and "Monocular 3D Pose and Shape Estimation of Multiple People in Natural Scenes - The Importance of Multiple Scene Constraints" which are incorporated in their entirety and are part of the present disclosure.

Human Appearance Transfer


Abstract



[0106] We propose an automatic person-to-person appearance transfer model based on explicit parametric 3d human representations and learned, constrained deep translation network architectures for photographic image synthesis. Given a single source image and a single target image, each corresponding to different human subjects, wearing different clothing and in different poses, our goal is to photo-realistically transfer the appearance from the source image onto the target image while preserving the target shape and clothing segmentation layout. Our solution to this new problem is formulated in terms of a computational pipeline that combines (1) 3d human pose and body shape estimation from monocular images, (2) identifying 3d surface colors elements (mesh triangles) visible in both images, that can be transferred directly using barycentric procedures, and (3) predicting surface appearance missing in the first image but visible in the second one using deep learning-based image synthesis techniques. Our model achieves promising results as supported by α perceptual user study where the participants rated around 65% of our results as good, very good or perfect, as well in automated tests (Inception scores and a Faster-RCNN human detector responding very similarly to real and model generated images). We further show how the proposed architecture can be profiled to automatically generate images of a person dressed with different clothing transferred from a person in another image, opening paths for applications in entertainment and photo-editing (e.g. embodying and posing as friends or famous actors), the fashion industry, or affordable online shopping of clothing.

1. Introduction



[0107] 

People are of central interest in images and video, so understanding and capturing their pose and appearance from visual data is critically important. While problems like deor 2d pose estimation have received considerable attention and witnessed significant progress recently, appearance modeling has been less explored comparatively, especially for bodies and clothing, in contrast to faces. One setback is that people are extremely sensitive to invalid human appearance variations and immediately spot them. This is to a large extent true for faces, as people are sharply tuned to fine social signals expressed as subtle facial expressions, but also stands true for human body poses, shapes and clothing. This makes it difficult to capture and possibly re-synthesize human appearance in ways that pass the high bar of human perception. While the realistic 3d human shape and appearance generation, including clothing, has been a long standing goal in computer graphics, with impressive studio results that occasionally pass the Turing test, these usually require complex models with sophisticated layering, manual interaction, and many cameras, which makes them difficult to use at large scale. For this purpose, flexible methods, that can be learned from data and can synthesize realistic human appearance are of obvious value. Arguably, even more important would be methods that can be controlled by image evidence in some way. For instance one may not just aim to generate plausible human shape and appearance in isolation - hard as this may be - but also condition on specific elements of pose and appearance in a given image in order to synthesize new ones based on it.

[0108] In this paper we formulate a new problem called human appearance transfer. Given a single source and a single target image of a person, each with different appearance, possibly different body shape and pose, the goal is to transfer the appearance of the person in the first image into the one of the person of the target image while preserving the target clothing and body layout. The problem is challenging as people are in different poses and may have different body shapes. A purely image warping or image to image translation approach would not easily generalize due to the large number of degrees of freedom involved in the transformation, e.g. the effect of articulation, depth and body shape on appearance. We provide a first solution that relies on fitting state-of-the-art 3d human pose and body models to both the source and the target images, transferring appearance using barycentric methods for commonly visible vertices, and learning to color the remaining ones using deep image synthesis techniques with appropriately structured 2d and 3d inputs. Example images, perceptual user studies, Inception scores [26], and the response of a state-of-the-art person detector confirm that the generated images of humans are perceptually plausible.

2. Related Work



[0109] Our work relies on 2d human detection and body part labeling [2, 24, 9], 3d pose estimation [24, 1, 29], parametric 3d human shape modeling [5, 30, 20, 18], procedures devoted to the semantic segmentation of clothing [28, 6, 17, 8], as well as image translation and synthesis methods [11, 3, 21, 14, 34, 3, 31].

[0110] Modeling the human appearance is a vast topic that has been approached on several fronts. One is through modifications of real images [10, 22], although the results are not entirely realistic. Computer graphics pipelines are also used, either in a mixed reality setting - where a moderately realistic graphics model is rendered in a real scene in a geometrically correct manner [10] - or e.g. by fitting a SCAPE model to real images [4]. In the former, the graphics character is still not photo-realistic; in the latter, clothing geometry is lacking. Detailed, accurate human shape estimation from clothed 3d scan sequences [33] can produce very good results but the acquisition setup is considerably more involved. Models to realistically capture complex human appearance including clothing in a laboratory setup, based on multiple cameras and relatively simple backgrounds appear in [15]. Procedures directed to the realistic acquisition of clothing exist [23, 33], but rely on an existing set of 3d models of garments and a 3d scanning device.

[0111] The methodology reviewed in the previous paragraph achieves excellent results under the application constraints it was designed for. However, some requires manual interaction, multiple cameras, simple backgrounds, specialized scanners, or complex modeling setups. In contrast, we aim at automatic appearance modeling in situations where one has no control on the acquisition setup and is given a minimal number of images (one or two). The idea is to exploit precise, but inherently limited in coverage, geometric estimation methods for the human pose and shape, and complement them with learning techniques, in order to achieve photo-realistic appearance transfer for specific images. There is relatively little research focusing on human appearance generation based on a combination of geometric and learning methods. One notable exception is the recent work by [13] which is able to generate realistic images of people given their silhouette, pose and clothing segmentation. The method relies on a variational auto-encoder [12] and a GAN [7, 11] for realistic image generation. However, it is not obvious how this model would perform inference for the appearance given an image, or how can it condition on a particular appearance and photo-realistically transfer it to a new pose. The human appearance transfer between two monocular images falls out of the domain of applicability of models like [13], and is the new problem defined and confronted in this research.

3. Human Appearance Transfer



[0112] Given a pair of RGB images - source and target, denoted by Is and It, each containing a person -, the main objective of our work is to transfer the appearance of the person from Is into the body configuration of the person from It, resulting in a new image Is⇒t.1 Our proposed pipeline is shown in fig. 2 and details are given in the next sections.
1The procedure is symmetric, as we can transfer in both directions.

3.1. 3D Human Pose and Body Shape Fitting



[0113] Human Detection & Body Parts Segmentation. To detect each person and infer critical semantic and geometric information, each image is fed through the Deep Multi-task Human Sensing (DMHS) network [24], a state-of-the-art predictor for body part labeling (semantic segmentation) and 3d pose estimation. DMHS is a multi-stage architecture, in which each stage refines the output from the previous stage, producing a tuple (J, B, R), where

is the set of 2d body joint configurations,

is the body part labeling map, and

is the 3d body joint configuration of the person detected in an image.

[0114] 3d Body Shape Fitting. We use the prediction of DMHS with the fitting method of [32] in order to estimate the human 3d body shape and pose from an image. The representation is based on the 3d SMPL body model [18]. A commonly used pipeline for fitting the model [1] relies on minimizing a sparse cost error over detected 2d human joint locations. However, the method of [32] utilizes all information available in the tuple (J, B, R) and the image. The 3d pose is initialized using R and refined so that each body part of the model aligns with the corresponding semantic segmentation labels in the image, based on DMHS estimates J and B (fig. 3). Please notice the difference between our body shape fitting procedure and the one of [1], illustrated in fig. 4. The head and arm orientations of our estimate are closer to the perceived one, due to a superior DMHS initialization (as opposed to a canonical T-pose) and the use of dense body parts semantic image segmentation labels during fitting.





[0115] The estimated 3d body model consists of a fixed number of Nv = 6890 vertices and a set of Nf = 13776 faces, F

that forms the triangle mesh. We define the fitted 3d body shape model as a tuple S = (C, D, M), where

encodes the RGB color at each vertex position,

encodes the disparity map of the fitted 3d mesh with respect to the image and M ∈ {0,1}Nv encodes the visibility of each vertex. Given the source and target images Is and It, we obtain human pose and shape estimates as mesh structures Ss and St, respectively.


3.2. Body Color Completion



[0116] We are first interested in estimating the pixel colors for the projected visible surface of St, denoted as Is→t, using the pixel colors on the projected visible surface of Ss.

[0117] Barycentric transfer. We begin by defining the common set of visible vertices ΛsΛt = {i|Ms(i) = 1 Λ Mt(i) = 1,1 ≤ iNv} and select the corresponding mesh faces FsΛt). For each face fFsΛt), we project it on Is and Is→t. For each pixel location in the projection of f on Is, we find its corresponding pixel location in the projection on Is→t using barycentric triangle coordinates. Finally, we copy the color information from one location to another.

[0118] Vertex Color Completion Network. The remaining set of visible vertices in St, Λt\s = {i|Mt(i) = 1 Λ Ms(i) = 0,1 ≤ iNv} needs to be colored. We rely on learning the implicit correlations among various body parts in order to propagate appearance information from the already colored vertex set Cs to Λt\s. Such correlations, effectively forms of pattern completion, are learned automatically from training data using a neural network.

[0119] Learning for Mesh Color Completion. We are given as inputs the color set Cs, the visibility mask Ms ∈ {0,1}Nv×Nv, and a binary mask that encodes the vertices we wish to color MΛt\s ∈ {0,1}Nv×Nv, i.e. visibility values are replicated along columns. The output is represented by the predicted colors

We define two weight matrices

and

Our network optimizes over these two matrices with the loss L defined as the Euclidean distance between the prediction

and the ground-truth colors CΛt\s :







where the softmax function is applied column-wise. Intuitively, any visible target vertex, without color, will have it assigned to the weighted mean (softmax function) of all the available colored vertex set, with weights encoded in matrices W1 and W2. We interpolate the predicted vertex colors

from the learned model over the corresponding mesh faces Ft\s), project the faces, and obtain the missing regions in Is→t.

[0120] Generating Training Samples. The training data consists of inputs, each being a subset of colored vertices from a mesh, and outputs that represent different subsets of vertices from the same mesh. In practice, given any monocular image, once we fit the 3d model, we can generate any possible input-output split over the visible vertices. However, our inputs tend to be structured, consisting of subsets of vertices seen in the source and target mesh as well as their difference set. To ensure a similar distribution, we take the inputs and outputs to be sets of visible vertices in the intersection of the source and target mesh (we assume intersection is non-trivial), and choose outputs in their difference sets, respectively. Two training examples can thus be generated, symmetrically, for every pair of images of people, with different appearance and in a different pose.

[0121] The drawback of this procedure is that, at training time, the network has to predict colors from a smaller intersection set of colored vertices (i.e. ΛsΛt), whereas at test time, it can use the fully colored set of vertices from the source mesh Cs.

3.3. Clothing Layout Warping



[0122] We use the model of [6] to estimate the clothing layout for target and source images, Lt and Ls, defined over a set of 20 clothing labels. Given the clothing layout source Ls, we want to transform it into the pose of the target image, Ls⇒t. We start by collecting clothing label information for each visible vertex in the source mesh Ss. We propagate the labeling on the entire mesh by using a geodesic nearest-neighbor approach. For labels not on the source





mesh, we collect the nearest-neighbour source vertex projections, vote on a displacement and translate the labeling accordingly. Thus, we obtain a rough estimate of Ls⇒t, which will denote by Ls→t. We gather additional image data from the web, consisting of source-target image pairs depicting the same person wearing the same clothes, but in different poses. On this dataset of ∼ 1,500 training pairs, we train an image to image translation network which outputs Ls⇒t given Ls→t and the disparity map Dt.

3.4. Human Appearance Synthesis



[0123] The previously described prediction Is→t captures the appearance transfer only as covered by our 3d body models. Hence, clothing layers (e.g. skirt, jacket) or hair which fall outside the coverage of the human body model are not transferred during the process. To achieve a higher perceptual quality for the generated image, we further refine our prediction using a Human Appearance Synthesis (HAS) network adapted based on ideas in [3]. This method performs multi-resolution refinement and was originally used in synthesizing photographic images conditioned on semantic segmentation layouts. Instead, we train the network to predict an image It given three types of inputs: a predicted semantic layout of clothing Lt or Ls⇒t, the disparity map Dt, and Is→t. The output of this HAS network, Is⇒t, represents our final, refined result.

4. Experiments



[0124] For all of our experiments we use the Chictopia10k dataset [16]. The images in this dataset depict different people, under both full and partial viewing, captured frontally. The high variability in color, clothing, illumination and pose makes this dataset suitable for our task. There are 17, 706 images available together with additional ground truth clothing segmentations. We do not use the clothing labels provided, but only the figure-ground segmentation such that we can generate training images cropped on the human silhouette.

[0125] We split the data in two subsets: 15, 404 images for training and 2,302 images for testing. We additionally prune some of the images based on the quality of body shape fitting, This is done by applying a soft threshold on the intersection over union (IoU) between the projection of the fitted body model and the foreground mask of the person. For each image in the dataset, we randomly select two other images from its corresponding subset (i.e. train or test) to construct image pairs. In the end, we use 28, 808 training pairs and 4, 080 testing pairs.



[0126] Appearance Transfer. Results of our Body Color Completion module are shown in fig. 9. Sample results of our Human Appearance Synthesis module are also given in fig. 10. Although we show transfer results in one direction, our method is symmetrical, so we obtain results of similar quality both ways, as shown in fig. 1.

[0127] Impact of Components and Failure Modes. Our Human Appearance Synthesis network receives as inputs Is→t, the depth map and the clothing segmentation of the target image. To evaluate the contribution of each of these inputs in the visual quality of the output, we train two additional Human Appearance Synthesis networks under similar conditions, but with different input data: one without the depth map, and the other without both the depth map and the clothing segmentation. In fig. 11, we provide visual results for all three networks. We observe that the best quality is obtained when using the complete network.

[0128] Errors occur in our pipeline when the clothing segmentation fails or the 3d body shape fitting does not yield good alignment with the person in the image. Examples are shown in fig. 12.


4.1. Identity Preserving Appearance Transfer



[0129] We also implement a variation of our model in order to preserve the identity of the target subject during appearance



transfer. To do so, we redefine the sets of vertices used in the Body Color Completion module. We start by identifying the skin and clothing image pixels of the target and source images by using the set of labels provided by the clothing segmentation model. We place the set of labels defined by hair, face, arms and legs in the skin category and the remaining ones in the clothing category. Then we assign a category (i.e. skin/cloth) for each vertex in the body models by inspecting the clothing labeling under their projections in the image. We fix the colors for the pixels/vertices categorized as skin in the target, and perform barycentric transfer only for the intersection of source and target vertices categorized as clothing. The colors for the remaining vertices are predicted as before by our Vertex Color Completion network. Sample results are shown in fig. 8.

4.2. Perceptual User Study and Automated Tests



[0130] We perform a perceptual study by asking 20 human subjects to evaluate our results. We present each one with 100 results in the form of source image (Is), target image (It) and our automatically generated appearance transfer image Is⇒t. We ask subjects to evaluate the appearance transfer quality of Is⇒t, by assigning it one of the following scores: very poor (1), poor (2), good (3), very good (4), perfect (5). Finally, we quantified their scores in the form of a normalized histogram, shown in fig. 13 (bottom). The mean score is 2.9, with standard deviation 0.96, suggesting that our transfer is reasonable on average.

[0131] We compare our method against recent work [19, 27] independently addressing the problem of pose conditioned human image generation. Such methods rely solely on information in the image without explicitly inferring 3d body pose. Set aside significant methodological differences, we are additionally able to perform identity-preserving transfer (fig. 8). Our results are more visually pleasing and superior in terms of Inception Scores [26], which are 3.09 [19], 3.35 [27] and 4.13 (Ours).



[0132] In order to understand possible difference in terms of image statistics, we also perform an automated test using a state-of-the-art human detector, Faster R-CNN [25]. We compute the human detection scores on two sets containing 2,000 generated and real images, respectively. In fig. 13 (top) we observe that the two distributions of detection scores are similar, with a dominant mode around value 0.99.

5. Conclusions



[0133] Modeling and synthesizing human appearance is difficult due to variability in human body proportions, shape, clothing and poses. However, models that can realistically synthesize complete images of humans under a degree of control (conditioning) on an input image appearance or pose, could be valuable for entertainment, photo-editing, or affordable online shopping of clothing. In this context, we define a new problem entitled human appearance transfer where given two images, source and target, of different peo with different poses and clothing, we learn to transfer the appearance of the source person on the body layout of the target person. Our solution relies on state-of-the-art 3d human pose and shape estimation based on deep multitask neural networks and parametric human shape modeling, combined with deep photographic synthesis networks controlled by appropriate 2d and 3d inputs. Our image results, backed-up by a perceptual user study, Inception scores, and the response of a state-of-the-art human person detector indicate that the proposed model can automatically generate images of humans of good perceptual quality, and with similar statistics as real human images. We also show how the model can be modified to realistically 'dress' a person shown in one image with clothing captured from a person in another image.


References



[0134] 

[1] F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, 2016.

[2] Z. Cao, T. Simon, S. Wei, and Y. Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, July 2017.

[3] Q. Chen and V. Koltun. Photographic image synthesis with cascaded refinement networks. In ICCV, October 2017.

[4] W. Chen, H. Wang, Y. Li, H. Su, Z. Wang, C. Tu, D. Lischinski, D. Cohen-Or, and B. Chen. Synthesizing training images for boosting human 3d pose estimation. In 3D Vision (3DV), 2016 Fourth International Conference on, pages 479-488. IEEE, 2016.

[5] R. Goldenthal, D. Hannon, R. Fattal, M. Bercovier, and E. Grinspun. Efficient simulation of inextensible cloth. ACM Transactions on Graphics (TOG), 26(3):49, 2007.

[6] K. Gong, X. Liang, X. Shen, and L. Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In CVPR, July 2017.

[7] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In NIPS, 2014.

[8] M. Hadi Kiapour, X. Han, S. Lazebnik, A. C. Berg, and T. L. Berg. Where to buy it: Matching street clothing photos in online shops. In ICCV, December 2015.

[9] K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask r-cnn. In ICCV, 2017.

[10] C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu. Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments. PAMI, 2014.

[11] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017.

[12] T. D. Kulkarni, W. Whitney, P. Kohli, and J. B. Tenenbaum. Deep convolutional inverse graphics network. In NIPS, 2015.

[13] C. Lassner, G. Pons-Moll, and P. V. Gehler. A generative model of people in clothing. In ICCV, 2017.

[14] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017.

[15] V. Leroy, J.-S. Franco, and E. Boyer. Multi-view dynamic shape refinement using local temporal integration. In ICCV, 2017.

[16] X. Liang, C. Xu, X. Shen, J. Yang, S. Liu, J. Tang, L. Lin, and S. Yan. Human parsing with contextualized convolutional neural network. In ICCV, 2015.

[17] Z. Liu, P. Luo, S. Qiu, X. Wang, and X. Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, pages 1096-1104, 2016.

[18] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. Smpl: A skinned multi-person linear model. SIGGRAPH, 34(6):248, 2015.

[19] L. Ma, X. Jia, Q. Sun, B. Schiele, T. Tuytelaars, and L. Van Gool. Pose guided person image generation. In NIPS, 2017.

[20] R. Narain, A. Samii, and J. F. O'Brien. Adaptive anisotropic remeshing for cloth simulation. ACM transactions on graphics(TOG), 31(6):152, 2012.

[21] A. Nguyen, J. Yosinski, Y. Bengio, A. Dosovitskiy, and J. Clune. Plug & play generative networks: Conditional iterative generation of images in latent space. In CVPR, 2017.

[22] L. Pishchulin, A. Jain, M. Andriluka, T. Thormaehlen, and B. Schiele. Articulated people detection and pose estimation: Reshaping the future. In CVPR, June 2012.

[23] G. Pons-Moll, S. Pujades, S. Hu, and M. Black. ClothCap: Seamless 4D clothing capture and retargeting. ACM Transactions on Graphics, (Proc. SIGGRAPH), 36(4), 2017. Two first authors contributed equally.

[24] A. Popa, M. Zanfir, and C. Sminchisescu. Deep Multitask Architecture for Integrated 2D and 3D Human Sensing. In CVPR, July 2017.

[25] S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NIPS, pages 91-99, 2015.

[26] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, X. Chen, and X. Chen. Improved techniques for training gans. In NIPS, 2016.

[27] A. Siarohin, E. Sangineto, S. Lathuiliere, and N. Sebe. Deformable gans for pose-based human image generation. In CVPR, 2018.

[28] E. Simo-Serra, S. Fidler, F. Moreno-Noguer, and R. Urtasun. A high performance crf model for clothes parsing. In ACCV, 2014.

[29] A. Toshev and C. Szegedy. Deeppose: Human pose estimation via deep neural networks. In CVPR, 2014.

[30] F. Xu, Y. Liu, C. Stoll, J. Tompkin, G. Bharaj, Q. Dai, H.-P. Seidel, J. Kautz, and C. Theobalt. Video-based characters: Creating new human performances from a multi-view video database. In ACM SIGGRAPH 2011 Papers, SIGGRAPH, pages 32:1-32:10, New York, NY, USA, 2011. ACM.

[31] C. Yang, X. Lu, Z. Lin, E. Shechtman, O. Wang, and H. Li. High-resolution image inpainting using multi-scale neural patch synthesis. In CVPR, 2017.

[32] A. Zanfir, E. Marinoiu, and C. Sminchisescu. Monocular 3D Pose and Shape Estimation of Multiple People in Natural Scenes - The Importance of Multiple Scene Constraints. In CVPR, 2018.

[33] C. Zhang, S. Pujades, M. Black, and G. Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In CVPR, 2017.

[34] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017.


Monocular 3D Pose and Shape Estimation of Multiple People in Natural Scenes The Importance of Multiple Scene Constraints


Abstract



[0135] Human sensing has greatly benefited from recent advances in deep learning, parametric human modeling, and large scale 2d and 3d datasets. However, existing 3d models make strong assumptions about the scene, considering either a single person per image, full views of the person, a simple background or many cameras. In this paper, we leverage state-of-the-art deep multi-task neural networks and parametric human and scene modeling, towards a fully automatic monocular visual sensing system for multiple interacting people, which (i) infers the 2d and 3d pose and shape of multiple people from a single image, relying on detailed semantic representations at both model and image level, to guide a combined optimization with feedforward and feedback components, (ii) automatically integrates scene constraints including ground plane support and simultaneous volume occupancy by multiple people, and (iii) extends the single image model to video by optimally solving the temporal person assignment problem and imposing coherent temporal pose and motion reconstructions while preserving image alignment fidelity. We perform experiments on both single and multi-person datasets, and systematically evaluate each component of the model, showing improved performance and extensive multiple human sensing capability. We also apply our method to images with multiple people, severe occlusions and diverse backgrounds captured in challenging natural scenes, and obtain results of good perceptual quality.

1. Introduction



[0136] Accurately detecting and reconstructing multiple people, possibly involved in interactions with each other and with the scene, based on images and video data, has extensive applications in areas as diverse as human-computer interaction, human behavioral modeling, assisted therapy, monitoring

sports performances, protection and security, special effects, modeling and indexing archival footage, or self-driving cars.

[0137] To support the level of modeling accuracy required by such applications, we ultimately need highly-detailed models able not just to detect people and their body joints in images, but also the spatial extent of body parts, as well as the three-dimensional pose, shape and motion for each person in the scene. For complex scenes, such demands would likely require a virtuous cycle between 2d and 3d reasoning, with feedback. One should further consider integrating anthropometry constraints, avoiding geometric collisions between the estimated models of multiple people, and reasoning about ground planes implicit in many scenes, as people rarely float, unsupported in space - and if so, usually not for long. Reconstructions must also be temporally fluid and humanly plausible. Most importantly, constraints need to be enforced in the context of an image observation process which - even with many cameras pointed at the scene - remains incomplete and uncertain, especially in scenarios where multiple people interact. While the integration of such constraints appears challenging, their use provides the opportunity to considerably restrict the degrees of freedom of any natural human parameterization towards plausible solutions.

[0138] In this paper, we address the monocular inference problem for multiple interacting people, by providing a model for 2d and 3d pose and shape reconstruction over time. Our contributions include (i) a semantic feedforward-feedback module that combines 2d human joint detection, semantic segmentation, and 3d pose prediction of people, with pose and shape refinement based on a novel semantic cost that aligns the model body parts with their corresponding semantic images regions, producing solutions that explain the complete person layout while taking into account its estimation uncertainty, (ii) incorporation of scene consistency measures including automatic estimation and integration of ground plane constraints, as well as adaptively avoiding simultaneous volume occupancy by several people, and (iii) resolution of the temporal person assignment problem based on body shape, appearance and motion cues within a Hungarian matching method, then solving a joint multiple-person smoothing problem under both 2d projection and 3d pose temporal fluidity constraints. Our quantitative results on datasets like Panoptic [12] and Human3.6M [11] validate the importance of the ingredients in the proposed design. Qualitative results in complex monocular images and video show that the model is able to reconstruct multiple interacting people in challenging scenes in a perceptually plausible way. The model also supports the realistic synthesis of human clothing and appearance (human appearance transfer) as shown in our companion paper [39].

2. Related Work



[0139] Our work relates to recently developed deep architectures for 2d human pose estimation [4, 9, 21, 35, 36], 3d human pose estimation based on fitting volumetric models [2, 15], feedforward deep models for 3d prediction [18, 22, 40], as well as integrated deep models for 2d and 3d reasoning [23, 27, 34, 19]. Accurate shape and motion-capture systems, based on multiple cameras or simplified backgrounds, have also been proposed with impressive reconstruction results [3, 7, 13, 26]. Systems designed for the 3d reconstruction of multiple people are relatively rare and existing ones are based on multiple cameras [1, 6, 5, 12, 14]. In [6], the method uses an arguably low number of cameras (3-4) to reconstruct several people, with promising results, but the level of interaction is somewhat limited. The work of [12] proposes a multi-person tracking system (which we also use for our 'ground-truth' monocular evaluation), although the system relies on a massive number of RGB and RGB-D cameras for inference, and the capture dome offers inherently limited background variability. Our single person initialization relies on the Deep Multitask Human Sensing Network (DMHS) [23] for initial 2d and 3d pose inference (body joints, semantic segmentation, pose prediction), which is then refined based on our own implementation of the human body model SMPL [15], augmented with learned semantic vertex labeling information, and using a new semantic loss function, which represents one of our contributions. Systems based on discriminative-generative (feedforward-feedback) components for 3d human pose estimation date, in principle, back at least to [25, 28, 31] but our approach leverages considerably different image representations, body models, cost functions and optimization techniques. Our automatic ground plane and adaptive people volume occupancy exclusion constraints, as well as our multiple people assignment and smoothing costs are integrated in a novel and coherent way, although monocular single person costs based on simpler model formulations and/or multiple hypotheses tracking techniques exist in the literature [2, 20, 24, 29, 30. 38].

3. Multiple Persons in the Scene Model



[0140] Problem formulation. Without loss of generality, we consider Np uniquely detected persons in a video with Nf frames. Our objective is to infer the best pose state variables

shape parameters

and individual person translations

with pNp and fNf. We start by first writing a per-frame, person-centric objective function



where the cost LS takes into account the visual evidence computed in every frame in the form of semantic body part labeling, LC penalizes simultaneous (3d) volume occupancy between different people in the scene, and LG incorporates the constraint that some of the people in the scene may have a common supporting plane. The term

is a Gaussian mixture prior similar to [2]. The image cost for multiple people under all constraints can be written as

If a monocular video is available, the static cost Lf is augmented with a trajectory model applicable to each person

once the temporal assignment throughout the entire video has been resolved. The complete video loss writes

where LT can incorporate prior knowledge on human motion, ranging from smoothness, assumptions of constant velocity or acceleration, or more sophisticated models learned from human motion capture data. In the next sections, we describe each cost function in detail.1
1Whenever unambiguous, we drop the f and p super-scripts.

[0141] In order to infer the pose and 3d position of multiple people we rely on a parametric human representation, SMPL [15], with a state-of-the-art deep multitask neural network for human sensing, DMHS [23]. In practice, we cannot assume a constant number of people throughout a video and we first infer the parameters B,Θ, T independently for each frame by minimizing the sum of the first two cost functions: LS and LC. Then, we temporally track the persons obtained in each frame by means of optimally solving an assignment problem, then re-optimize the objective, by adding the temporal and ground plane constraints, LT and LG. For those people detected in only some of the frames, optimization will proceed accordingly over the corresponding subset. An overview of the method is shown in fig. 2.

3.1. Single Person Feedforward-Feedback Model



[0142] SMPL [15] is a differentiable parametric human model - represented by template vertices V0 - and controlled by pose vectors

and shape parameters

The pose of the model is defined by a standard skeletal rig that has the main body joints. For each body part, the vectors controlling the pose are provided in axis-angle representations of the relative rotations w.r.t. their parents in the kinematic tree. The axis angle for every joint is transformed to a rotation matrix using the Rodrigues transformation. The shape parameters β impact limb size, height and weight and represent coefficients of a low dimensional shape space learned from registered meshes. SMPL provides matrix functions dependent on θ and β, namely

which gives the transformed vertex positions for the whole mesh, and

which outputs the joint positions for the associated kinematic tree. The total number of vertices in the SMPL model is nV = 6890 and the total number of joints in the kinematic tree is nJ = 24. For simplicity of explanation, let v denote V(θ,β|V0) and let x be J(θ,β|V0). We refer to the translation of the model in camera space as



[0143] DMHS [23] is a state-of-the-art feedforward multi-task deep neural network for human sensing that provides, for a given image

the following estimates: the 2d and 3d joints of a single person as well as the semantic body parts at pixel level. We denote these 3 outputs by the matrices



and

respectively. We denote by mJ = 17 the number of joints in the representation considered by the network and Ns = 25 the number of semantic body parts. The method has been shown to perform well for both indoor images as well as outdoor. The challenges of integrating DMHS and SMPL stem from accurately fitting (transferring) the parametric SMPL model to the 3d joint positions predicted by DMHS, as well as designing semantic-based cost functions that allow to efficiently couple the model to the observations - perform 3d fitting in order to best explain the human layout in the image. In order to semantically assign model mesh components to corresponding image regions, one needs a consistent 'coloring' of their vertices according to the NS human body part labels available e.g. in Human3.6M [11]. This can be achieved robustly, during a training process. We project and fit the SMPL model in multiple (4) views and for different ground truth poses from Human3.6M (we chose 100 different poses). Then each model vertex was associated the median image body part label, available in Human3.6M, transferred from images to the corresponding

vertex projections. See fig. 4 for coloring examples.

3.1.1 Feedforward Prediction, Pose & Shape Transfer



[0144] We detail the transfer procedure for a single person and perform the same steps for all people in each frame of a video. To transfer the feedforward prediction of the configuration of joints y3D obtained from DMHS to the SMPL model, we have to define a cost function Φ3d(θ,β), and infer optimal θ and β parameters. One such cost function is the Euclidean distance between joints shared in both representations (i.e. i,jCJ, where 1 ≤ imJ and 1 ≤ jnJ and CJ is the set of compatible joint indices)

where h indicates the index of the pelvis and x(j) - x(h) represents the centered 3d pose configuration with respect to the pelvis joint. Unless otherwise stated, we use ∥·∥ for the ℓ2 norm, ∥·∥2.

[0145] However, based on (4) the DMHS to SMPL transfer is unsatisfactory. This is because 1) the prediction made by DMHS is not necessarily a valid human shape, and 2) a configuration in the parameter space of β or even in the space of θ does not necessarily represent an anatomically correct human pose. In [2], multiple regularizers were added: a norm penalty on β and a prior distribution on θ. However, these risk excessive bias.

[0146] We propose an alternative transfer equation, focusing on qualitatively modeling the pose predicted by DMHS so to preserve the 3d orientation of limbs. Our function Φcos penalizes the cosine distance between limbs - or selected pairs of joints - that are shared in both representations (property denoted by (i,j), (a,b) ∈ CL where 1 ≤ i,jmJ and 1 ≤ k,lnJ). Given aij = y3D(i) - y3D(j) and bkl = x(k) - x(l), the cost is



[0147] While in practice the minimization of Φcos converges quickly to a perfect solution (often close to 0) and the resulting pose is perceptually similar to DMHS, the implicit shape information provided by DMHS is lost. In situations where the original 3d joint prediction confidence is high (e.g. training and testing distributions are expected to be similar, as in Human3.6M), one can further optimize over β, starting from solutions of (5)



Results of the proposed transfer variants are shown in fig. 3.

3.1.2 Semantic 3d Pose and Shape Feedback



[0148] 



[0149] After transferring the pose from DMHS to SMPL we obtain an initial set of parameters θ0 and β0 and one can refine the initial DMHS estimate. One way to fit the 3d pose and shape model starting from an initialization [2, 32], is to minimize the projection error between the model joints and the corresponding detected image joint locations, y2d. We denote by

the image projection function, with fixed camera intrinsincs. One possible loss is the Euclidean distance, computed over sparse joint sets weighted by their detection confidence w (some may not be visible at all)



[0150] The problem of minimizing θ and β for monocular error functions, defined over distances between sparse sets of joints, is its ambiguity, as the system is clearly under-determined, especially for depth related state space directions that couple along camera's ray of sight [33]. We propose a new error function based on projecting the mesh v in the image I and measuring the dense, pixel-wise semantic error between the semantic segmentation transferred by the model projection and a given DMHS semantic body part segmentation prediction yS.

[0151] We are given NS semantic classes that describe body parts with yS storing semantic confidence maps. We construct a function fS(p = (x,y)T) = argmaxkyS(p,k) with 1 ≤ xW,1 ≤ yH integers, that returns the body part label 1 ≤ kNS of pixel location p in the image I. Let vk be vertices pertaining to the body part indexed in k and

their image projection.

[0152] We design a cost ΦS(Θ, B, T), where each point p from the semantic body part segmentation maps finds its nearestneighbour in pk, and drags it in place. Appropriately using pixel label confidences (x,y) for a given class k as yS is important for robust estimates in a cost that writes



[0153] In practice, our semantic cost is further weighted by a normalization factor 1/Z, with

ensuring φS remains stable to scale transformations impacting the area of the semantic map (closer or further away, with larger or smaller number of pixels, respectively). Another desirable property of the semantic loss is that when confidences are small, ΦS will have a lower weight in the total loss, emphasizing other qualitatively different terms in the cost. The total semantic loss can then be written


3.2. Simultaneous Volume Occupancy Exclusion



[0154] To ensure that estimated models of people in a scene are not inconsistent, by occupying the same 3d space volume simultaneously, we need additional processing. We design adaptive representations to first compute enclosing parallelepipeds for each person according to its current model estimates, rapidly test for intersections (far-range check), and only integrate detailed, close range collision avoidance into the loss when the far-range response is negative. For close-range volume occupancy exclusion, we use specialized terms obtained as follows: for each person model, we fit tapered superquadrics to each limb, and represent the limb by a series of Nb fitted spheres inside the superquadric, with centers c and radius r. For any two persons, p and p', we define the loss LC(p,p') based on distances between all spheres belonging, respectively, to the first and second person





[0155] The loss for Np persons in a frame is defined as the sum over all pair-wise close-range losses LC(p,p') among people with negative far-range tests. People with positive far-range tests do not contribute to the volume occupancy loss. Notice how this cost potentially couples parameters from all people and requires access to their estimates. See fig. 5 for visual illustrations.

3.3. Ground Plane Estimation and Constraint



[0156] We include a prior that the scene has a ground-plane on which, on average, the subjects stand and perform actions. To build a correct hypothesis for the location and orientation of the plane, we design a cost that models interactions between the plane and all human subjects, but leaves room for outliers, including people who, temporarily or permanently, are not in contact with the ground. Specifically, we select the 3d ankle positions of all persons in all the frames of a video, be these xi, and fit a plane to their locations.

[0157] We assume that a point z is on a plane with a surface normal n if the following equation is satisfied (z - p)Tn = 0, where p is any fixed point on the plane. Given that some of the ankles might be occluded, we use a confidence term to describe the impact they have on the fitting process. We use the confidence wi from the DMHS 2d joint detector, with a two-folded purpose to 1) select the initial point p belonging to the plane as the weighted median of all ankle locations of the detected persons, and 2) weight measurements used in the robust L1 norm estimate of the plane hypothesis. Our plane estimation objective is





[0158] The estimates (p, n*) are then used in the ground-plane constraint term LG to penalize configurations with 3d ankle joints estimates away from the plane

Where the subscripts l and r identify the left and the right ankles for a person p at time f. The weighting of the terms is performed adaptively based on confidences wl, wr of the associated ankle joints. If these are not visible, or are visible within some distance of the ground and not confident, constraints are applied. If the joints are visible and confident, or far from the ground, constraints are not applied.

3.4. Assignment and Trajectory Optimization



[0159] Independently performing 3d human body pose and shape optimization in a monocular video can lead to large translation variations along depth directions and movements that lack natural smoothness. For this reason, we propose a temporal constraint that ensures for each of the inferred models that estimates in adjacent frames are smooth. To achieve it, we first need to resolve the assignment problem over time (identify or track the same individual throughout the video), then perform temporal smoothing for each individual track.

[0160] To solve the person assignment problem, we use the Hungarian algorithm to optimally build tracks based on an interframe inter-person cost combining the appearance consistency (measured as distances between vectors containing the median colors of the different body parts, computed over the model vertices), the body shape similarity, and the distance between the appropriately translated 3d joints inferred for each person, at frame level.

[0161] Once the assignment has been resolved between every pair of estimated person models in every successive set of frames, and tracks are built, several motion priors can be used, ranging from a constant velocity model, to more sophisticated auto-regressive processes or deep recursive predictors learned from training data [17, 8, 37]. The integration of such motion representations in our framework is straightforward as long as they remain differentiable. Here we experiment with constant velocity priors on pose angles, Θ as well as translation variables, T. Our temporal loss function component at

frame f ≥ 2 for a person (track) p is defined as



[0162] The shape parameters βp are set as the median of

f. Because smoothing axis-angle representations is difficult, the angle-related costs in (15) are represented using quaternions, which are easier to smooth. Gradients are propagated through the axis-angle to quaternion transformation during the optimization.

4. Experiments



[0163] We numerically test our inference method on two datasets, CMU Panoptic [12] and Human3.6M [11], as well as qualitatively on challenging natural scenes (see fig. 7). On Human3.6M we test different components of the model including semantic feedback, smoothing and the effect of multiview constraints. Panoptic in turn provides the real quantitative test-bed for the complete monocular system.

[0164] Given a video with multiple people, we first detect the persons in each frame and obtain initial feedforward DMHS estimates for their 2d body joints, semantic segmentation and 3d pose. Similarly to [16], we extend DMHS to partially visible people, by fine-tuning both the semantic and the 3d pose estimation components of DMHS on a partial view version of Human80K[10]. For each person we perform the transfer proposed in (5) that aligns the limb directions of 3d estimates predicted by DMSH with the limb directions of SMPL. The transfer gives an initialization for pose and shape. The initial translation of each person is set to 3 meters in front of the camera.
Human3.6M is a large-scale dataset that contains single person images recorded in a laboratory setup using a motion capture system. The dataset has been captured using 4 synchronized RGB cameras and contains videos of 11 actors performing different daily activities. We select 3 of the most difficult actions: sitting, sitting down and walking dog to test
Table 1: Automatic 3d human pose and translation estimation errors (in mm) on the Panoptic dataset (9,600 frames, 21,404 people). Notice the value of each component and the impact of the ground-plane constraint on correct translation estimation.
 HagglingMafiaUltimatumPizzaMean
MethodPoseTranslationPoseTranslationPoseTranslationPoseTranslationPoseTranslation
DMHS [23] 217.9 - 187.3 - 193.6 - 221.3 - 203.4 -
2d Loss 135.1 282.3 174.5 502.2 143.6 357.6 177.8 419.3 157.7 390.3
Semantic Loss 144.3 260.5 179.0 459.8 160.7 376.6 178.6 413.6 165.6 377.6
Smoothing 141.4 260.3 173.6 454.9 155.2 368.0 173.1 403.0 160.8 371.7
Smoothing Ground Plane 140.0 257.8 165.9 409.5 150.7 301.1 156.0 294.0 153.4 315.5
Table 2: Mean per joint 3d position error (in mm) on the Human3.6M dataset, evaluated on the test set of several very challenging actions. Notice the importance of various constraints in improving estimation error.
MethodWalkingDogSittingSitting Down
DMHS [23] 78 119 106
Semantic Loss 75 109 101
Multi View 51 71 65
Smoothing 48 68 64
our single-person model. We use the official left-out test set from the selected actions, consisting of 160K examples. On this dataset we can only evaluate the pose inference under MPJPE error, but without the translation relative to the camera. We show results in table 2. We obtain an improvement over DMHS by using the proposed semantic 3d pose and shape feedback, cf. (10). On this dataset, we also experiment with multi-view inference and show a consistent improvement in 3d pose estimation. For multi-view inference, the loss function proposed in (10) is easily extended as a sum over measurements in all available cameras. Adding a temporal smoothness constraint further reduces the error. We also evaluated our method on all 15 actions from the official test set (911,744 configurations) and obtain an average error of 69 mm.2
2Detailed results can be seen at http://vision.imar.ro/ human3.6m/ranking.php (TestsetH36M_NOS10).

[0165] CMU Panoptic Dataset. We selected data from 4 activities (Haggling, Mafia, Ultimatum and Pizza) which contain multiple people interacting with each other. For each activity we selected 2 sub-sequences, each lasting 20 seconds (i.e. 600 frames), from HD cameras indices 30 and 163. In total, we obtain 9,600 frames that contain 21,404 people. We do not validate/train any part of our method on this data.
3For variability only, all testing is monocular.

[0166] Evaluation Procedure. We evaluate both the inferred pose, centered in its hip joint, under mean per joint position error (MPJPE), and the estimated translation for each person under standard Euclidean distance. We perform the evaluation for each frame in a sequence, and average the results across persons and frames. We match each ground-truth person in the scene with an estimation of our model. For every ground-truth pose, we select the closest inferred model under the Euclidean distance, in camera space.

[0167] Ablation Studies. We systematically test the main components of the proposed monocular inference system and show the results detailed for each activity in table 1. Compared to DMHS, our complete method reduces the MPJPE error significantly, from 203.4 mm to 153.4 mm on average (-25%), while also computing the translation of each person in the scene. The translation error is, on average, 315.5 mm. The semantic projection term helps disambiguate the 3d position of persons and reduces the translation error compared to using only the 2d projection term. Temporally smoothing the pose estimates decreases the translation error further. Imposing the ground plane constraint makes the most significant contribution in this setup, decreasing the total translation error from 371 mm to 315 mm (-15%). Even though the total pose error also decreases when all constraints are imposed, on some sequences (e.g. Haggling) the error did not decrease when semantic terms are used. At a closer look, we noticed that the semantic maps and 3d initialization from DMHS were particularly noisy on those sequences of Haggling, camera index 30. Qualitative results in monocular images from the Panoptic dataset are shown in fig. 6. Our method produces perceptually plausible 3d reconstructions with good image alignment in scenes with many people, some only partially visible, and captured under non-conventional viewing angles.

5. Conclusions



[0168] We have presented a monocular model for the integrated 2d and 3d pose and shape estimation of multiple people, under multiple scene constraints. The model relies on feedforward predictors for initialization and semantic fitting for feedback and precise refinement (shape adaption) to the observed person layout. It estimates and further integrates ground plane and volume occupancy constraints, as well as temporal priors for consistent, plausible estimates, within a single joint optimization problem over the combined representation of multiple people, in space and time. Our experimental evaluation, including ablation studies, is extensive, covers both single-person and multiple-person datasets and illustrates the importance of integrating multiple constraints.



Moreover, we qualitatively show that the method produces 3d reconstructions with tight image alignment and good perceptual quality, in both monocular images and video filmed in complex scenes, with multiple people, severe occlusion and challenging backgrounds. To our knowledge, such a large-scale fully automatic monocular system for multiple person sensing under scene constraints has been presented here for the first time.

References



[0169] 

[1] V. Belagiannis, S. Amin, M. Andriluka, B. Schiele, N. Navab, and S. Ilic. 3d pictorial structures for multiple human pose estimation. In CVPR, 2014.

[2] F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, and M. J. Black. Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image. In ECCV, 2016.

[3] A. Boukhayma, J.-S. Franco, and E. Boyer. Surface motion capture transfer with Gaussian process regression. In CVPR, 2017.

[4] Z. Cao, T. Simon, S. Wei, and Y. Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017.

[5] A. Elhayek, E. Aguiar, A. Jain, J. Tompson, L. Pishchulin, M. Andriluka, C. Bregler, B. Schiele, and C. Theobalt. Efficient ConvNet-based marker-less motion capture in general scenes with a low number of cameras. In CVPR, 2015.

[6] A. Elhayek, E. de Aguiar, A. Jain, J. Thompson, L. Pishchulin, M. Andriluka, C. Bregler, B. Schiele, and C. Theobalt. Marconi-convnet-based marker-less motion capture in outdoor and indoor scenes. IEEE transactions on pattern analysis and machine intelligence, 39(3):501-514, 2017.

[7] D. A. Forsyth, O. Arikan, L. Ikemoto, J. O'Brien, and D. Ramanan. Computational Studies of Human Motion: Tracking and Motion Synthesis. NOW Publishers Inc, 2006.

[8] K. Fragkiadaki, S. Levine, P. Felsen, and J. Malik. Recurrent network models for human dynamics. In ICCV, pages 4346-4354, 2015.

[9] E. Insafutdinov, L. Pishchulin, B. Andres, M. Andriluka, and B. Schiele. DeeperCut: A deeper, stronger, and faster multi-person pose estimation model. In ECCV, 2016.

[10] C. Ionescu, J. Carreira, and C. Sminchisescu. Iterated second-order label sensitive pooling for 3d human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1661-1668, 2014.

[11] C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3d human sensing in natural environments. PAMI, 2014.

[12] H. Joo, H. Liu, L. Tan, L. Gui, B. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Y. Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, 2015.

[13] V. Leroy, J.-S. Franco, and E. Boyer. Multi-view dynamic shape refinement using local temporal integration. In ICCV, 2017.

[14] Y. Liu, J. Gall, C. Stoll, Q. Dai, H.-P. Seidel, and C. Theobalt. Markerless motion capture of multiple characters using multiview image segmentation. IEEE transactions on pattern analysis and machine intelligence, 35(11):2720-2735, 2013.

[15] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. SMPL: A skinned multi-person linear model. SIG-GRAPH, 34(6):248:1-16, 2015.

E. Marinoiu, M. Zanfir, V. Olaru, and C. Sminchisescu. 3D Human Sensing, Action and Emotion Recognition in Robot Assisted Therapy of Children with Autism. In CVPR, 2018.

J. Martinez, M. J. Black, and J. Romero. On human motion prediction using recurrent neural networks. CoRR, abs/1705.02445, 2017.

J. Martinez, R. Hossain, J. Romero, and J. J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017.

D. Mehta, S. Sridhar, O. Sotnychenko, H. Rhodin, M. Shafiei, H.-P. Seidel, W. Xu, D. Casas, and C. Theobalt. Vnect: Realtime 3d human pose estimation with a single rgb camera. ACM Transactions on Graphics (TOG), 36(4):44, 2017.

R. M. Neal, M. J. Beal, and S. T. Roweis. Inferring state sequences for non-linear systems with embedded hidden Markov models. In NIPS, 2004.

G. Papandreou, T. Zhu, N. Kanazawa, A. Toshev, J. Tompson, C. Bregler, and K. Murphy. Towards accurate multi-person pose estimation in the wild. In CVPR, 2017.

G. Pavlakos, X. Zhou, K. G. Derpanis, and K. Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In CVPR, 2017.

A. Popa, M. Zanfir, and C. Sminchisescu. Deep multitask architecture for integrated 2d and 3d human sensing. In CVPR, 2017.

V. Ramakrishna, T. Kanade, and Y. Sheikh. Reconstructing 3d human pose from 2d image landmarks. ECCV, 2012.

D. Ramanan and C. Sminchisescu. Training deformable models for localization. In CVPR, 2006.

H. Rhodin, N. Robertini, D. Casas, C. Richardt, H.-P. Seidel, and C. Theobalt. General automatic human shape and motion capture using volumetric contour cues. In ECCV, 2016.

G. Rogez and C. Schmid. Mocap-guided data augmentation for 3d pose estimation in the wild. In NIPS, 2016.

L. Sigal, A. Balan, and M. J. Black. Combined discriminative and generative articulated pose and non-rigid shape estimation. In NIPS, 2007.

E. Simo-Serra, A. Quattoni, C. Torras, and F. Moreno-Noguer. A joint model for 2d and 3d pose estimation from a single image. In CVPR, 2013.

C. Sminchisescu and A. Jepson. Variational mixture smoothing for non-linear dynamical systems. In CVPR, 2004.

C. Sminchisescu, A. Kanaujia, and D. Metaxas. Learning joint top-down and bottom-up processes for 3d visual inference. In CVPR, 2006.

C. Sminchisescu and B. Triggs. Estimating Articulated Human Motion with Covariance Scaled Sampling. IJRR, 22(6):371-393, 2003.

C. Sminchisescu and B. Triggs. Kinematic jump processes for monocular 3d human tracking. In CVPR, 2003.

B. Tekin, P. Marquez Neila, M. Salzmann, and P. Fua. Learning to fuse 2d and 3d image cues for monocular body pose estimation. In ICCV, 2017.

J. J. Tompson, A. Jain, Y. LeCun, and C. Bregler. Joint training of a convolutional network and a graphical model for human pose estimation. In NIPS, 2014.

A. Toshev and C. Szegedy. Deeppose: Human pose estimation via deep neural networks. In CVPR, 2014.

J. M. Wang, D. J. Fleet, and A. Hertzmann. Gaussian process dynamical models. In NIPS, 2006.

H. Yasin, U. Iqbal, B. Kruger, A. Weber, and J. Gall. A dual-source approach for 3d pose estimation from a single image. In CVPR, 2016.

M. Zanfir, A. Popa, and C. Sminchisescu. Human appearance transfer. In CVPR, 2018.

X. Zhou, M. Zhu, K. Derpanis, and K. Daniilidis. Sparseness meets deepness: 3d human pose estimation from monocular video. In CVPR, 2016.




Claims

1. A method, comprising:

determining in a first image a person and a first 3D human pose and body shape fitting model, wherein the person has a first pose and a first clothing;

determining in a second image a person and a second 3D human pose and body shape fitting model, wherein the person has a second pose and a second clothing; and

generating, by use of the first 3D human pose and body shape fitting model and by use of the second 3D human pose and body shape fitting model, an image comprising the person of the first image in the first pose and with the second clothing and/or generating, by use of the first 3D human pose and body shape fitting model and by use of the second 3D human pose and body shape fitting model, an image comprising the person of the second image in the second pose and with the first clothing.


 
2. The method according to claim 1, wherein the first 3D human pose and body shape fitting model comprises a first number of vertices and a first set of faces that form a first triangle mesh, and wherein the second 3D human pose and body shape fitting model comprises a second number of vertices and a second set of faces that form a second triangle mesh.
 
3. The method according to claim 2, wherein the first 3D human pose and body shape fitting model represents a tuple St = (Ct, Dt, Mt), wherein, with regard to the first triangle mesh, Ct encodes a RGB color at each vertex position, Dt encodes a disparity 3d depth map of the first triangle mesh with respect to the first image, and Mt encodes a visibility of each vertex, and wherein the second 3D human pose and body shape fitting model represents a tuple Ss = (Cs, Ds, Ms), wherein, with regard to the second triangle mesh, Cs encodes a RGB color at each vertex position, Ds encodes a disparity map of the first triangle mesh with respect to the first image, and Ms encodes a visibility of each vertex.
 
4. The method according to any one of the preceding claims, wherein the first 3D human pose and body shape fitting model is obtained by determining, with regard to the person in the first image and by using a multitask deep neural network model, corresponding 2D body joint locations, semantic body part segments and 3D pose and by refining the first 3D human pose and body shape fitting model by executing non-linear optimization, and wherein the second 3D human pose and body shape fitting module is obtained by determining, with regard to the person in the second image and by using the multitask deep neural network model, corresponding 2D body joint locations, semantic body part segments and 3D pose and by refining the second 3D human pose and body shape fitting model by executing non-linear optimization.
 
5. The method according to claim 4, wherein the refining the first 3D human pose and body shape fitting model by executing non-linear optimization comprises aligning, with regard to the person in the first image, a corresponding articulated human body mesh with a semantic segmentation layout obtained by using the multitask deep neural network model with regard to the person in the first image, and wherein the refining the second 3D human pose and body shape fitting model by executing non-linear optimization comprises aligning, with regard to the person in the second image, a corresponding articulated human body mesh with a semantic segmentation layout obtained by using the multitask deep neural network model with regard to the person in the second image.
 
6. The method according to claim 4 or 5, wherein the non-linear optimization is performed by executing semantic fitting where vertices of the respective mesh model carry a body type which is matched to the corresponding body type pixels detected in the image in the sense that vertices of a particular type of 3d model should onto pixels detected of the same type in the image.
 
7. The method according to any one of the preceding claims 2 to 6, wherein the generating of the image comprises:

determining a common set of visible vertices visible in the first triangle mesh with regard to the person in the first image and in the second triangle mesh with regard to the person in the second image;

determining a divergent set of vertices visible in the first triangle mesh with regard to the person in the first image and not visible in the second triangle mesh with regard to the person in the second image;

assigning colors to the vertices in the common set of visible vertices by using a barycentric transfer;

training a first neural network module for body color completion for assigning colors to vertices of the first triangle mesh that are visible and that are not present in the common set of visible vertices; and

training a second neural network module for human appearance synthesis for assigning colors of resulting appearance image of the first person, that are part of clothing, but cannot be modeled by the mesh model of the first person and the first neural network for body color completion.


 
8. The method according to claim 7, wherein the training of the first neural network module comprises using weakly-supervised learning based on sampling subsets of visible vertices of a 3d mesh fitted to images of people, as source and target set for self-training, respectively.
 
9. The method according to claim 7 or 8, wherein the second neural network module for human appearance synthesis is trained to predict its final output, given among inputs, clothing segmentation layout, disparity map and results predicted by the first neural network for body color completion.
 
10. The method according to one of claims 7 to 9, wherein the first and the second neural network modules are trained in an end-to-end process and using a multi-task loss.
 
11. The method of any one of the preceding claims, wherein the generating of the image comprises transferring a clothing layout of the second image to a clothing layout of the first image.
 
12. A device configured to execute a method according to any one of the preceding claims 1 to 11.
 
13. A system configured to execute a method according to any one of the claims 1 to 11.
 
14. A computer program comprising instructions which when executed by a processor carry out the method of one of the claims 1 to 11.
 




Drawing



































































Search report


















Search report




Cited references

REFERENCES CITED IN THE DESCRIPTION



This list of references cited by the applicant is for the reader's convenience only. It does not form part of the European patent document. Even though great care has been taken in compiling the references, errors or omissions cannot be excluded and the EPO disclaims all liability in this regard.

Non-patent literature cited in the description