(19)
(11)EP 3 608 844 A1

(12)EUROPEAN PATENT APPLICATION

(43)Date of publication:
12.02.2020 Bulletin 2020/07

(21)Application number: 18306104.3

(22)Date of filing:  10.08.2018
(51)Int. Cl.: 
G06N 3/04  (2006.01)
G06N 3/08  (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)Applicant: Naver Corporation
Seongnam-si, Gyeonggi-do 13561 (KR)

(72)Inventor:
  • WEINZAEPFEL, Philippe
    38330 Montbonnot-Saint-Martin (FR)

(74)Representative: Regimbeau 
20, rue de Chazelles
75847 Paris Cedex 17
75847 Paris Cedex 17 (FR)

  


(54)METHODS FOR TRAINING A CRNN AND FOR SEMANTIC SEGMENTATION OF AN INPUTTED VIDEO USING SAID CRNN


(57) The present invention relates to a method for training a convolutional recurrent neural network, CRNN, for semantic segmentation in videos; the method being characterized in that it comprises the implementation, by a data processor (11a) of a first server (1a), of steps of:
(a) Training from a base of training images already semantically segmented, a first convolutional neural network, CNN;
(b) Training from a base of training videos already semantically segmented, a recurrent convolutional neural network, CRNN, corresponding to the first CNN wherein a convolutional layer has been replaced by a recurrent module having a hidden state; said training comprising, for each pair of successive frames (t - 1, t ∈ 1; T 2) of a video of said base of training videos already semantically segmented:
(b1) Warping the internal state of the recurrent layer according to an estimated optical flow between the frames of the pair, so as to adapt the internal state to the motion of pixels between the frames of the pair;
(b2) learning parameters of at least the recurrent module


A method for semantic segmentation of an inputted video is further proposed.












Description

FIELD OF THE INVENTION



[0001] The field of this invention is that of machine learning.

[0002] More particularly, the invention relates to methods for training a recurrent convolutional neural network for semantic segmentation of videos, and for using such recurrent convolutional neural network.

BACKGROUND OF THE INVENTION



[0003] Computer vision is a recent field of endeavor which aims at providing computers with high-level understanding from digital videos. It seeks to automate tasks that the human visual system can do.

[0004] One of these tasks is the so-called "semantic" segmentation. Generally speaking, segmentation is the process of partitioning an image into sets of pixels. When each of these sets corresponds to an entity whose type can be identified (a car, a person, a building, etc.), the segmentation is considered semantic. In practice, semantic segmentation consists in assigning a class label to each pixel among a predetermined set of classes.

[0005] It has a strong interest in many applications including robotics (to understand the scene and identify where the robot can navigate) and autonomous driving. In a similar spirit, semantic segmentation is useful in the context of augmented reality to understand the scene, find the extents of objects and of regions in which the virtual objects can be added.

[0006] Semantic segmentation is a challenging task in computer vision due to the large variety of classes that include stuffs (sky, grass, etc.) and well-defined objects (cars, bottles, etc.) that can be large (e.g. building) or thin (e.g. traffic light). Semantic segmentation outputs are smooth both in space (neighborhood pixels tend to belong to the same class, except at object boundaries) and in time (a real-world point has constant label in time, which means its projections also).

[0007] It has been proposed to use Convolutional Neural Networks (CNNs) to this end. They are a type of Neural Networks wherein the connection pattern between the neurons is inspired by the visual cortex of animals. They are thus especially suitable for video processing, they indeed allow efficiently the recognition of entities in images.

[0008] Thus, after a supervised learning stage wherein a CNN is trained by feeding it a training set of videos already segmented, i.e. provided with the labels of to each pixel in the frames of the videos, the CNN becomes able to segment on its own some inputted videos (which are not labelled, in particular "fresh" videos from live CCTV).

[0009] Most existing methods to semantic segmentation performs at the image level, i.e. an algorithm is applied in every frame of the video independently.

[0010] The main idea of the prior art is to transform an image classification CNN architecture into a fully convolutional version (see for instance the document Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. (2015)) in order to output a dense prediction. Note that feature maps usually have lower resolution than the input images, due to the stride of standard computer vision CNN architectures. The first deep segmentation approaches were thus refining the output using graphical models such as conditional random fields (CRFs) using an upsampling strategy, i.e., an encoder-decoder architecture.

[0011] More recently, dilated convolutions (also called atrous convolutions) have been introduced (see for instance the document Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. In: ICLR. (2016)) and allow one to extract denser feature maps from existing classification CNN architectures.

[0012] For better accuracy, a recent trend consists in modeling the context information at multiple scales in the last layers. For instance, this is done in the Atrous Spatial Pyramid Pooling (ASPP) of DeepLab (see the document Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. PAMI (2017)) by concatening the output of dilated convolutions at multiple factors, as well as globally pooled image features in Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. Technical Report (2017), or by performing spatial pooling at several grid scales in PSP-Net (see for instance Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: CVPR. (2017)).

[0013] Typically, this leads to a dissatisfying flickering effect: some regions of the real world undergo many changes of semantic labels between consecutive frames. This is illustrated in Figure 1b which shows examples of per-frame estimation on three consecutive frames (figure 1a). We observe noise artifacts in the predicted labels across time (boxes), with even a region that contains flickering between more than 2 labels (right box).

[0014] To improve the consistency across time, the semantic segmentation should be performed as the video level, i.e. on the frames together.

[0015] There have only been a few recent attempts to address semantic video segmentation. First approaches were based on a post-processing stage using graphical models such as Markov Random Fields, Perturb-and-MAP Random Fields or CRFs in space and time. More recently, is has been proposed in the document Kundu, A., Vineet, V., Koltun, V.: Feature space optimization for semantic video segmentation. In: CVPR. (2016), to leverage optical flow to model motion in the pairwise potentials between frames. Another method to refine semantic segmentation in videos consists in using a filtering strategy. All these approaches nevertheless do not produce consistent video segmentation outputs.

[0016] More recently, it was proposed in the document Gadde, R., Jampani, V., Gehler, P.V.: Semantic video CNNs through representation warping. In: ICCV. (2017) to introduce the NetWarp module to integrate some temporal consistency into semantic segmentation CNN architectures. The idea is to combine the features of the current frame with the features from the previous frame warped according to the optical flow. In practice, features from consecutive frames are aggregated after warping them according to the optical flow and are used to make the final prediction, but they remain limited to a predefined and fixed number of frames.

[0017] Alternatively, the document Shelhamer, E., Rakelly, K., Hoffman, J., Darrell, T.: Clockwork convnets for video semantic segmentation. In: ECCV. (2016), proposes the Clockwork CNN architecture for semantic video segmentation. The idea consists in reusing intermediate features from previous frames, with the aim of reducing the runtime of video segmentation however at the cost of a drop of accuracy.

[0018] There is a consequently need for an improvement in segmentation methods using neural networks which could allow to directly leverage temporal information and outperform any known technique, without any theoretical restriction.

SUMMARY OF THE INVENTION



[0019] For these purposes, the present invention provides according to a first aspect a method for training a convolutional recurrent neural network, CRNN, for semantic segmentation in videos;
the method being characterized in that it comprises the implementation, by a data processor of a first server, of steps of:
  1. (a) Training from a base of training images already semantically segmented, a first convolutional neural network, CNN;
  2. (b) Training from a base of training videos already semantically segmented, a recurrent convolutional neural network, CRNN, corresponding to the first CNN wherein a convolutional layer has been replaced by a recurrent module having a hidden state; said training comprising, for each pair of successive frames of a video of said base of training videos already semantically segmented:

    (b1) Warping the internal state of the recurrent layer according to an estimated optical flow between the frames of the pair, so as to adapt the internal state to the motion of pixels between the frames of the pair;

    (b2) learning parameters of at least the recurrent module.



[0020] As it will be explained, providing a standard convolutional recurrent layer in which the internal state is warped between frames according to the optical flow (which can be estimated using another CNN) sensibly improve semantic segmentation in videos, especially in term of consistency across time.

[0021] Preferred but non limiting features of the present invention are as follow:
  • the pair of successive frames comprises a previous frame and a current frame, the optical flow estimated being the backward optical flow from the current frame to the previous frame, sub step (a1) comprising applying a warping function to the internal state associated to the previous frame so as to obtain a warped internal state associated to the previous frame corresponding to the internal state wherein each pixel has underwent a displacement according to said backward optical flow;
  • the warping function is such that ∀pt,ht-1(pt)=ht-1(pt+Ftt-1(pt)), wherein pt is a pixel of the current frame t, Ftt-1 is the backward optical flow, ht-1 is the internal state associated to the previous frame t - 1, ht-1 is the warped internal state associated to the previous frame t - 1;
  • step (b) comprises a previous sub-step (b0) of estimating the optical flow between the frames of the pair, performed using a second CNN, step (a) comprising training from a base of training pairs of successive frames between which the optical flow is known, the second CNN;
  • said second CNN is a FlowNetSimple network or a FlowNetCorrelation network;
  • the parameters of each layer of the CRNN before the recurrent module are fixed during sub-step (b2);
  • the convolutional layer of the first CNN which is replaced by a recurrent module is the penultimate convolutional layer;
  • the parameters of the CRNN learnt at sub-step (b2) are those of the recurrent module and the last convolutional layer of the first CNN;
  • sub-step (b2) comprises further learning the parameters of the second CNN;
  • the first CNN comprises before said penultimate convolutional layer an Atrous Spatial Pyramid Pooling module;
  • the recurrent module is chosen among a convolutional Gated Recurrent Unit and a convolutional Long Short-Term Memory.


[0022] In a second aspect, the invention provides a method for semantic segmentation of an inputted video, characterized in that it comprises the implementation of steps of:

(a') Training, by data processing means of a first server, from at least a base of training videos already semantically segmented, a recurrent convolutional neural network, CRNN, comprising a recurrent module having a hidden state;

(b') Performing, by processing means of a second server, for each pair of successive frames of the inputted video:

(b'0) estimating the optical flow between the frames of the pair;

(b'1) Warping the internal state of the recurrent layer according to the estimated optical flow, so as to adapt the internal state to the motion of pixels between the frames of the pair;

(b'2) Semantically segmenting each frame with the CRNN.



[0023] Preferred but non limiting features of the present invention are as follow:
  • the step (a') is implemented in accordance with the learning method according to the first aspect.
  • the step (a') comprises training from a base of training pairs of successive frames between which the optical flow is known, a convolutional neural network, CNN, sub-step (b'0) being performed using said CNN.


[0024] According to a third and a fourth aspects, the invention provides a computer program product, comprising code instructions to execute a method according to the first aspect for training a convolutional recurrent neural network, CRNN, or a method according to the second aspect for semantic segmentation of an inputted video; and a computer-readable medium, on which is stored a computer program product comprising code instructions for executing a method according to the first aspect for training a convolutional recurrent neural network, CRNN, or a method according to the second aspect for semantic segmentation of an inputted video.

BRIEF DESCRIPTION OF THE DRAWINGS



[0025] The above and other objects, features and advantages of this invention will be apparent in the following detailed description of an illustrative embodiment thereof, which is to be read in connection with the accompanying drawings wherein:
  • figure 1a represents an example of three successive frames;
  • figure 1b represents semantic segmentation of the example frames of figure 1a using a known per-frame technique;
  • figure 1c represents semantic segmentation of the example frames of figure 1b using the method according to the invention;
  • figure 2 illustrates an example of architecture in which the method according to the invention is performed;
  • figure 3 illustrates how the optical flow displaces pixels over an example of two consecutive frames;
  • figure 4a represents an example of a first CNN architecture for use in the methods according to the invention;
  • figure 4b represents an example of a CRNN architecture for use in the methods according to the invention.

DETAILED DESCRIPTION OF A PREFERRED EMBODIMENT


Architecture



[0026] According to two complementary aspects of the present invention are proposed:
  • a method for training a Convolutional Recurrent Neural Network (CRNN, or simply Convolutional RNN) for semantic segmentation in videos;
  • a method for semantic segmentation of an inputted video using a CRNN, advantageously trained according to the first method.


[0027] Any video is a sequence of T frames (according to its length), numbered from 1 to T. In other words, a generic frame of a video will be referred to as frame t, with t

1; T

. Each frame is an image, i.e. a matrix of pixels of a given size, for instance 321 x 321 pixels.

[0028] By semantic segmentation in a video, it is meant as previously explained classification of each pixel of each frame of the video, i.e. for each frame the prediction of a label of each pixel, defining a type of an entity depicted by said pixel, among a predetermined list of labels. It is to be understood that the frame is thus partitioned into a plurality of sets of pixels with the same label, each "connected" set of pixels with the same label defining an entity, i.e. a "real word" object. For instance, all the pixels depicting one car should be labelled as pixels of the "car" type.

[0029] The above mentioned two types of methods are implemented within an architecture such as illustrated in Figure 2, by means of a first and/or second server 1a, 1b. The first server 1a is the learning server (implementing the first method) and the second server 1b is a segmentation, server (implementing the second method). It is fully possible that these two servers be merged.

[0030] Each of these servers 1a, 1b is typically remote computer equipment connected to an extended network 2 such as the Internet for data exchange. Each one comprises data processing means 11a, 11b of processor type (in particular the data processing means 11a of the first server have strong computing power, since learning is long and complex compared with ordinary use of the trained CRNN), and optionally storage means 12 such as a computer memory e.g. a hard disk.

[0031] The first server 1a stores or has access to one of more training database i.e. a set of already classified data (as opposed to so-called inputted data that precisely is sought to be classify). As it will be explained, the data is here constituted of images and/or videos, and at least comprises training videos already semantically segmented. In other words, pixels of the training videos are already labelled.

[0032] The architecture advantageously comprises one or more items of client equipment 10, which may be any work station (also connected to network 2), preferably separate from the servers 1a, 1b but possibly being merged with one and/or the other thereof. The client equipment 10 has one or more data items to be classified. The operators of the equipment are typically "clients" in the commercial meaning of the term, of the service provider operating the first and/or second servers 1a, 1b.

Convolutional RNN



[0033] Recurrent Neural Networks (RNNs) are a type of Neural Networks that allows one to leverage sequential information, sequences of frames in the present case. This means that the output at a given frame will not only depend on the features from this frame, but also on the predictions from previous frames, thanks to an "internal state", also called hidden state or memory. They are very suitable for tasks such as handwriting recognition, or speech recognition.

[0034] RNNs have proven effective for sequence modeling in neural networks: the idea is to learn an internal state that accumulates relevant information across time and to base the prediction on the current input and this internal state. RNNs are often hard to train due to vanishing gradient issues.

[0035] RNNs includes a "RNN module", which is a block of one of more layers presenting the "recurrent" behavior. RNN modules are well known to the skilled person, see for example Gated Recurrent Unit, GRU (Cho, K., Van Merri¨enboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y.: Learning phrase representations using rnn encoder-decoder for statistical machine translation. EMNLP (2014)) or Long Short-Term Memory, LTSM (Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation (1997)). GRU and LSTM contain learnable gates in order to selectively update the internal state, thus allowing for propagating the gradients throughout longer sequences while training.

[0036] Generally speaking, if defining a general setting in which xt, ht and ot are respectively the input, the hidden state and the output of a RNN module from the frame t. The output ot of the RNN module is a function of the features xt at the current frame t as well as the hidden state ht-1 at the previous frame t - 1. At the same time, the hidden state ht is updated based on the same inputs, so that a RNN module is defined by the formula:

h0 is initialized (h0) with a tensor of all zeros. The training stage at this zero initialization is preferably performed with sequences longer sequences than pairs.

[0037] The present methods propose to adapt RNN modules to the task of semantic segmentation so as to take into account the inherent motion in videos.

[0038] Indeed, xt and ot are presently feature maps for the task of semantic segmentation, so convolutional operators (in which the input, the internal state and the output are 3D tensors) could be used inside the RNN module, for instance a ConvGRU (Ballas, N., Yao, L., Pal, C., Courville, A.: Delving deeper into convolutional networks for learning video representations. ICLR (2016)) or a ConvLSTM (Xingjian, S., Chen, Z.,Wang, H., Yeung, D.Y., Wong,W.K.,Woo,W.c.: Convolutional Istm network: A machine learning approach for precipitation nowcasting. In: NIPS. (2015)).

[0039] These convolutional operators have been used for instance for video frame prediction and video captioning. Convolutional RNNs have also been applied to other dense video prediction tasks, such as video object segmentation in which the goal is to output the binary segmentation of an object across the video given its segmentation in the first frame.

[0040] It has been proposed in the document Siam, M., Valipour, S., Jagersand, M., Ray, N.: Convolutional gated recurrent networks for video segmentation. In: ICIP. (2017), to apply ConvGRU to semantic video segmentation. However, such models lack motion modeling: the prediction at a given pixel is based on the history of the prediction at this pixel, while this pixel might represent a different point in the real world in case of motion.

[0041] Indeed, the output ot shall depend on a local spatial neighborhood of xt and ht-1, with the size of the neighborhood depending on the kernel size of the convolutional operators in a known fashion.

[0042] Standard CRNNs could be contemplated, but only in cases where a given pixel pt at a frame t and the same pixel pt-1 at the previous frame t - 1 (by "same pixel", it is meant the pixel at the exact same coordinates) are projections of the same real-world point, i.e., if this point remains static in the real world (or moves along the projection ray). In other words, this would assume that the video is static, i.e., that there is no motion. However, in many applications, like autonomous driving, such an assumption is not valid and pixels may move significantly between frames.

[0043] More precisely, the prediction at the pixel pt of the frame t normally depends on the features xt in the local neighborhood

of this pixel pt as well as the local neighborhood

from the hidden state ht-1 at the previous frame t - 1. To the contrary, the prediction at pt should actually be based on the features xt in

as well as the local neighborhood

from the hidden state ht-1, where p't-1 is the pixel corresponding to the same point as pt in the real world.

[0044] Figure 3, which represents two examples of successive frames t - 1 (left) and t (right), shows the pixels pt-1 and qt-1 on the previous frame t - 1 at the exact same coordinates of two pixels pt and qt on the frame t. pt-1 and qt-1 do not belong to the same region as pt and qt respectively, in contrast to the pixels p't-1 and q't-1.

[0045] The so-called "optical flow" is a vector field defining the displacement of all pixels from an image to another one (in particular a pair of successive frames t - 1,t

1; T

2, i.e. a previous frame t - 1 and a current frame t). For instance, the optical flow Ft-1→t defines the displacement of all pixels from the previous frame t - 1 to the current frame t, and the optical flow Ftt-1 ("backward flow") defines the displacement of all pixels from the current frame t to the previous frame t - 1.

[0046] The pixel p't-1 is given by the formula p't-1 = pt + Ftt-1(pt): in other words, the optical flow "shifts" each pixel according to the vector field.

[0047] In order to adapt the feature maps to the motion of all pixels according to the flow, the present methods astutely propose combining the use of a recurrent module with the use of a "warping" function. Indeed, if at-1 is a feature map at previous frame t - 1 and at the corresponding feature map at current frame t (feature map being vector fields, i.e. associating a vector to each pixel), we would like to combine for any pixel features of at(pt) and at-1(p't-1) = at-1(pt + Ftt-1(pt)). Since this operation would be hard to implement efficiently (because the pixels involved have different coordinates), it is proposed to first compute the so-called warping of the features, i.e. to build the vector field at-1 defined as at-1(pt) = at-1(pt + Ftt-1(pt)). Thus, features of at(pt) and at-1(pt) can be directly combined pixel by pixel.

[0048] To rephrase, knowing the optical flow (Ftt-1), the warping is the computing from any given feature map at-1, of a warped feature map at-1 corresponding to the given feature map at-1 wherein each pixel has undergone a displacement according to the optical flow.

[0049] It is to be understood that the warping does not modify the values of the feature map (i.e. the vectors) but just "spatially re-arranges" them: in the present case of a warping from the previous frame t - 1 to the current frame t, the warping function associates to each pixel pt of the current image t the value of the feature map to be warped (the internal state ht-1 associated to the previous image t - 1) for another pixel p't-1 according to the backward optical flow Ftt-1.

[0050] In practice, the optical flow contains float values and not only integers. A warping function can be based on a bilinear interpolation (e.g. between 4 values), which is differentiable, except in the exceptional case where the flow values are integers, in which case the gradient is set to zero.

[0051] A generic warped RNN module, nicknamed "FlowingRNN" can thus be written as a RNN module in which the internal state is warped between frames according to the optical flow, i.e.:



[0052] In a preferred embodiment, is proposed a "FlowingGRU", i.e. a FlowingRNN based on the ConvGRU module. The following formula can thus be given:









[0053] Where ⊙ is the element-wise multiplication, ★ the convolution operator, σ the sigmoid function, ReLU the rectified linear unit non-linearity (see below), and W's and b's are learnable parameters (weights and biases respectively).

[0054] Intuitively, the reset gate rt is learning how to combine the input xt with the previous hidden state ht-1 while the update gate zt is learning how much of the previous memory should be kept.

[0055] The main difference between the present FlowingGRU and the standard ConvGRU is that the hidden state is warped according to the optical flow, and thus the prediction at a given pixel and timestep is based on the history of this particular point in the real world, even in the case of pixel motion. One minor difference is also that a ReLU non-linearity is preferably used instead of a standard tanh as ReLU is more common in computer vision CNN architectures.

[0056] The skilled person will be able to transpose the invention to any other type of recurrent module. In particular, the applicant has successfully tested a "FlowingLSTM", i.e. a FlowingRNN based on the ConvLSTM module.

[0057] Generally speaking, FlowingRNN can be plugged into any fully convolutional image segmentation approach so as to perform an enhanced semantic segmentation of videos, with increased performance, especially in term of consistency across time.

[0058] Figure 1c shows examples of semantic segmentation using FlowingRNN on the three consecutive frames of figure 1a. With respect to figure 1b, we observe an absence of noise artifacts in the predicted labels across time (boxes).

[0059] The convolutional RNN, the flow estimation and the warping modules are all differentiable, making it possible to train the full network end-to-end.

Training method



[0060] In a first aspect, there is proposed a training method, implemented by the data processing means 11a of the first server 1a. Said method trains the Convolutional Recurrent Neural Network, CRNN (or Convolutional RNN), for semantic segmentation in videos.

[0061] In a first step (a), a standard Convolutional Neural Network, CNN (not recurrent), which will be referred to as "first" CNN, is trained from a base of training images already semantically segmented (for example the MS-Coco dataset, see Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll'ar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. (2014)).

[0062] In other words, the first CNN is a classic "per-frame" baseline which does not take into account local dependency across time. Any known architecture can be used for the first CNN.

[0063] The CNNs generally contain four types of layers processing the information:
  • the convolution layer which processes blocks of the image one after the other;
  • the non-linear layer (also called correction layer) which allows pertinence of the result to be improved by applying an "activation function";
  • the pooling layer which allows the grouping together of several neurons in one single neuron;
  • the fully connected layer that connects all the neurons of a layer with all the neurons of the preceding layer.


[0064] It will be here noted that in case of a fully convolutional network, the fc layer(s) do(es) not exist anymore and are(is) transformed into a convolutional layer of kernel size 1. This allows to output a map of probabilities instead of just one probability distribution for the entire image.

[0065] Are known in the realm of CNNs batch normalization layers to accelerate learning of data.

[0066] At the present time, the activation function of the non-linear layer the most frequently used is the ReLU function (Rectified Linear Unit) which is equal to f(x) = max(0, x), and the pooling layer the most used is the MaxPool2×2 function which corresponds to a maximum between four values of a square (four values are pooled into one), but other polling functions such as AveragePool or SumPool.

[0067] The convolution layer, denoted CONV, and the fully connected layer denoted FC, generally correspond to a scalar product between the neurons of the preceding layer and the weights of the CNN.

[0068] Typical CNN architectures stack a few pairs of layers CONV → RELU then add a layer POOL and repeat this scheme [(CONV → RELU)p → POOL] until a sufficiently small output vector is obtained, and terminate with one or two fully connected layers FC.

[0069] In the example of figure 4a is proposed an architecture for the first CNN inspired by DeepLab with a ResNet-101 backbone (see the document He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. (2016)).

[0070] In this preferred architecture, the first convolution layer as well as block1 to block4 are from ResNet-101, except that block3 and block4 are using dilated convolutions with a factor (also called rate) of 2 and 4 respectively, reducing the stride of the network from 32 to 8. Next is an ASPP (Atrous Spatial Pyramid Pooling) module in which the results from different convolutional kernels with various dilation factors are stacked. The output of the ASPP module is fed to a 1x1 convolution layer (with batch normalization and ReLU) with 256 filters, before a final 1x1 convolution layer that outputs class scores. These scores are bilinearly upsampled to the initial image resolution, to apply for example the per-pixel softmax loss at training time, or to estimate the predicted class by taking the per-pixel argmax at test time.

[0071] To train the first CNN, a Stochastic Gradient Descent (SGD) with a batch size of 16 can be used. In particular, could be employed a so-called 'poly' learning rate schedule, i.e., the initial learning rate is multiplied by

at the iteration i over a total number of N iterations. Training patches come from random crops of the training images with size for example of 321x321 pixels, while the CNN can be tested on the full images. At training, data augmentation is preferably performed by randomly left-right flipping the images, and applying a random scaling factor in the range [0:5; 1:5], before taking the random crop.

[0072] Step (a) advantageously comprises the further training of a second CNN from a base of training pairs of successive frames between which the optical flow is known.

[0073] As it will be explained, the second CNN aims at estimating the optical flow in pairs of frames, in particular a backward optical flow Ftt-1 from a current frame t to the previous frame t - 1.

[0074] The second CNN may be such as FlowNetSimple or FlowNetCorrelation (see the document Dosovitskiy, A., Fischer, P., Ilg, E., Hausser, P., Hazirbas, C., Golkov, V., van der Smagt, P., Cremers, D., Brox, T.: Flownet: Learning optical flow with convolutional networks. In: ICCV. (2015)). The architecture FlowNetSimple (generally just called FlowNetS) is preferred and used in the example of figure 4b. In this architecture, both input images are stacked together and fed through a network only made of convolutional layers.

[0075] ReLU non-linearity and 'upconvolution' can also be implemented as convolutions.

[0076] Said base of training pairs of successive frames between which the optical flow is known is for example the FlyingChair dataset, or the Ground Truth Optical Flow Dataset.

[0077] In a second step (b), the CRNN is itself trained. The idea is to build the CRNN from the first CNN (and the second CNN if there is). Indeed, the CRNN corresponds to the first CNN wherein a convolutional layer has been replaced by a recurrent module having a hidden state as explained (a FlowingRNN module such as FlowingGRU).

[0078] Preferably, as represented by figure 4b, the convolutional layer of the first CNN which is replaced by a recurrent module is the penultimate convolutional layer. This position offers excellent results as it is just after the ASPP module in the example based on ResNet-101, but the recurrent module may be placed elsewhere in the first CNN, for instance replace the last convolutional layer. Also, there could be three convolutional layers after the ASPP module, and the recurrent module could replace the first one (i.e. be the antepenultimate convolutional layer of the first CNN).

[0079] The step (a) could be seen as an "initialization" of the parameters of the CRNN, and the step (b) a "finetuning" of these parameters by a further training from a base of training videos already semantically segmented (see below).

[0080] Some of parameters obtained thanks to the training of the first CNN can be fixed, i.e. not further learnt in step (b), preferably the parameters of each layer of the CRNN before the recurrent module (i.e. until ASPP included). The parameters further learnt are then preferably those of the recurrent module, the last convolutional layer of the CNN (and generally each convolutional layer not fixed), and possibly the second CNN (if there is).

[0081] Most popular benchmarks for semantic segmentation are limited to images and cannot be used for step (b). While the present method can be trained even if only a few pixels are annotated in the videos, by ignoring pixels that are not annotated in the loss, training with dense ground-truth would likely lead to better temporal consistency.

[0082] Most real-world datasets have only one frame annotated per sequence, while some synthetic benchmarks are limited by the realism of the data or by the low number of sequences, so that said base of training videos already semantically segmented is preferably the recent Viper dataset (see the document Richter, S.R., Hayder, Z., Koltun, V.: Playing for benchmarks. In: ICCV. (2017)) which consists of around 250k frames from almost 200 video sequences captured from the realistic Grand Theft Auto V video game. The synthetic rendering allows one to obtain ground-truth for multiple tasks, including semantic segmentation. Illumination and weather conditions vary from day to night, from sunny to rainy, snowy or foggy, making the dataset challenging. The semantic segmentation task has 23 classes including stuff (e.g. 'sky', 'terrain'), vehicles (e.g. 'car', 'truck') and small objects (e.g. 'traffic sign', 'traffic light').

[0083] The step (b) advantageously starts with a sub-step (b0) of estimating, for each pair of successive frames t - 1, t

1; T

2 of a video of said base of training videos already semantically segmented, the optical flow between the frames of the pair (in particular the backward optical flow Ftt-1 from the current frame t to the previous frame t - 1).

[0084] This step is preferably performed using the second CNN (FlowNetS in the figure 4b) using as input the both frames of the pair, but the present method is not limited to any technique for estimating the optical flow.

[0085] It is to be noted that the optical flow could be already estimated between frames of the training base so this step (b0) stays optional.

[0086] In a sub-step (b1), as already explained, the internal state of the recurrent layer is warped according to the estimated optical flow, so as to adapt the internal state to the motion of pixels between the frames of the pair.

[0087] In sub-step (b2) parameters of at least the recurrent module are learning.

[0088] In addition to the recurrent module, since the output of this layer will be different from the one of the per-frame baseline, the following layers are also re-learnt. As already explained, the parameters further learnt are then preferably those of the recurrent module, the last convolutional layer of the CNN (and generally each convolutional layer not fixed), and possibly the second CNN (if there is).

[0089] To this end, can be used SGD using backpropagation though time algorithm with a poly learning rate schedule similar as for the initial training of the first CNN, with a batch size of 4 and training sequences of 12 consecutive frames. Similar data augmentation strategy can be also used. At test time, the hidden state from the previous frame can be used, without limitation on the sequence length.

Classification method



[0090] In a second aspect, there is proposed an inputted video semantic segmentation method implemented by the data processing means 11b of the second server 1b. In other words, the method according to the second aspect performs the semantic segmentation of an inputted video, i.e. labels pixels of the frames of the inputted video.

[0091] Said inputted video to be segmented can be received from the client equipment 10.

[0092] At a first step (a'), the training of a CRNN is performed by the first server 1a (from at least a base of training videos already semantically segmented, and advantageously a base of training images already semantically segmented). Preferably, said training is consistent with the method according to the first aspect, i.e. a first CNN is trained, so that is constructed a CRNN based on the first CNN wherein the penultimate convolutional layer is replaced by a recurrent module of the "FlowingRNN" type.

[0093] Alternatively, or in combination, the step (a') may further comprise training from a base of training pairs of successive frames between which the optical flow is known, a convolutional neural network, CNN (i.e. the second CNN of the training method).

[0094] It is to be understand that either the first and the second server 1a, 1b are the same equipment, or step (a') comprise the transmission of the trained CRNN parameters and weights from the first server 1a to the second server 1b, for storage in the memory 12.

[0095] At a second step (b'), the processing means 11b of the second server 1b uses the CRNN for the segmentation.

[0096] To this end, step (b') comprises, for each pair of successive frames (t - 1, t

1; T

2) of the inputted video:

(b'0) estimating the optical flow between the frames of the pair;

(b'1) Warping the internal state of the recurrent layer according to the estimated optical flow, so as to adapt the internal state to the motion of pixels between the frames of the pair;

(b'2) Semantically segmenting each frame with the CRNN Said sub-steps (b'0) and (b'1) are advantageously equivalent to sub-steps (b0) and (b1) and can include any refinement thereof.



[0097] Finally, the classification result obtained can be transmitted back to said client equipment 10.

Tests



[0098] The efficiency of the present method could be evaluated using the standard per-pixel mean Intersection-over-Union metric (mIoUP): for each class, the ratio of the number of pixels correctly predicted for this class over the number of pixels whose ground-truth or prediction is this class is computed, and the average over all classes is reported.

[0099] However, this does not measure the temporal consistency of the semantic segmentation prediction. To this end, a mean Intersection-over-Union may be further computed at the trajectory level, in which a trajectory has to be consistently labeled across all frames with the ground-truth class in order to be considered as correct. dense trajectories over the full videos can be extracted (see for example the document Sundaram, N., Brox, T., Keutzer, K.: Dense point trajectories by gpu-accelerated large displacement optical flow. In: ECCV. (2010)), with optical flow extracted at half resolution and a density of 8 pixels. Only the trajectories for which the ground-truth labeling is consistent are kept in order to filter out noisy trajectories due to the errors in the flow estimation. In practice, this filtering stage removes about 8% of the extracted trajectories.

[0100] It finally allows reporting the "per-trajectory" mean Intersection over Union (mIoUT), in which trajectories whose pixels have different estimated labels are count as wrong. The consistency in each class could also computed as the following: among the trajectories whose ground-truth is consistently a given class, is computed the ratio of trajectories for which the prediction is consistent, whatever the predicted class. The "mean consistency" (mCons.) could be reported by averaging over all classes.

[0101] Firstly, the impact of the length T of the training sequences on the performance of the preferred FlowingGRU module is studied. At test time, the hidden state of the previous frame is used in all cases, independently of the length of the sequences used at training. As a sanity check, the model is firstly used with an all-zeroes internal state in all cases and sequences of 1 frame at training, so as to validate that the performance is similar to the per-frame baseline. T is then increased from 2 to 12. A consistent gain in performance is observed when T increases. The mIoUP is higher by 1.2%, the mIoUT by 2.1% and the mCons. by 3.5% when using training sequences of 12 frames compared to the per-frame baseline.

[0102] Then, the impact of the number of output channels C of the convolutions (and thus of the internal state also) in the preferred FlowingGRU module is studied, as well as the kernel size K (of all convolutions in FlowingGRU, see equations page 13). We observe optimal performances with C = 256 and K = 3, which can be explained by the facts that:
  • a lower number of channels does not suffice to carry relevant information, both spatially from the previous layer and over time from the previous frames for an accurate semantic segmentation;
  • a kernel with a size higher than 1 allows to integrate information over a local neighborhood, and thus to recover from small errors in the optical flow computation;
  • a kernel with a size higher than 3 has too many parameters to effectively learn local temporal consistency.


[0103] The impact of the proposed FlowingGRU module (with the preferred values T = 12, K = 3 and C = 256) is compared to some variants and to related works in Table 1 (left). Note that neither of these works have experimented on the Viper dataset and they use different backbone architectures. For a fair comparison, the same baseline is used, their respective proposed modules are plugged at the same place as FlowingRNN and the same training scheme is followed.
Table 1
 ViperCityscapes
mIoUPmIoUTmCons.mIoUP
per-frame 69.78 71.19 82.68 73.17
FlowingGRU 71.00 73.33 86.12 74.31
FlowingGRU (tanh) 70.18 71.68 83.64 73.11
FlowingGRU (fixed flow) 70.13 71.57 83.95 71.14
ConvGRU 70.24 71.70 83.57 70.70
ConvLSTM 65.62 68.64 81.10 71.77
FlowingLSTM 69.68 70.42 82.42 72.87
ConvGRU last 69.96 71.99 83.82 72.96
FlowingGRU last 70.40 72.90 85.84 74.02


[0104] The FlowingGRU module (with ReLU non-linearity) compared to the per-frame baseline shows a clear improvement of 1.22% in mIoUP, 2.14% in mIoUT and 3.44% in mCons. This highlights that FlowingGRU effectively leverages temporal information for dense video prediction tasks while taking into account pixel motion.

[0105] The FlowingGRU module with ReLU non-linearity could be compared to two variants. The first one consists in using a standard tanh nonlinearity in the equation page 13 (instead of ReLU), see row 'FlowingGRU (tanh)' in Table 1. We observe a drop of 0.8% in mIoUP, showing that ReLU is more suitable for semantic segmentation. In the second variant the parameters of FlowNetS are fixed, and not finetuned during training (i.e. the second CNN is not further learnt in step (b)), see row 'FlowingGRU (fixed flow)' in Table 1. In this case, the decrease is of 0.9% in mIoUP, showing the importance of training the flow estimation together with the FlowingGRU module.

[0106] The FlowingGRU module is further compared to other recurrent models. A standard ConvGRU is firstly tried, without warping of the internal state, as used for instance in the context of semantic video segmentation. Results are reported in row 'ConvGRU' of Table 1. The three metrics are slightly increased compared to the per-frame baseline, showing that it indeed learns some temporal consistency. However, ConvGRU performs significantly worse than the preferred FlowingGRU, with for instance a mIoUP 0.8% lower. Indeed, by keeping the internal state as it is between consecutive frames, ConvGRU assumes that the pixels at the same location in consecutive frames are projections of the same real-world point, which is wrong in most cases. Similarly, the mIoUT and the mCons. also significantly decrease by 1.63% and 2.55% respectively.

[0107] Next, GRU is replaced by LSTM, in which a ReLU non-linearity is also used instead of tanh, see rows 'ConvLSTM' and 'FlowingLSTM' in Table 1. The performance is significantly lower than with GRU. One explanation is that LSTMs are harder to train than GRUs for vision tasks. One interesting result in that FlowingLSTM performs significantly better than the Convolutional LSTM (+4.06% in mIoUP), highlighting again the importance of warping the internal state.

[0108] Finally, the position of the FlowingGRU module could be studied by replacing the last convolutional layer instead of the penultimate one. In this case, note that the output goes into a softmax (no non linearity could be used). We observe that this variant leads to a decrease of 0.6% mIoUP on the Viper dataset. When using ConvGRU, the performance is also lower when placing the module at the last layer. By replacing the penultimate layer, the history is embedded in a latent space from which the final prediction is made, which is more powerful than directly predicting the final segmentation from the hidden state and the current features.

[0109] Several experiments could also be run on the real-world Cityscapes dataset (Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. (2016)) using the fine annotation only, i.e., 2975 sequences for training and 500 for validation. Each sequence has 30 frames with annotation on 1 frame only (the twentieth frame).

[0110] The training and testing procedures can be kept similar to the case of the Viper dataset. The only difference is that ground-truth is annotated on sparse frames. Therefore, during training, the loss for frames that are not annotated is ignored, making video information less suitable to be leveraged. the noisy trajectories cannot be cleaned with only one annotated frame per sequence, so that only mIoUP may be reported.

[0111] Results for different variants are reported in the right column of Table 1. First, it can be observed that FlowingGRU outperforms the per-frame baseline with a gain of 1.14% in mIoUP. In all cases, the 'Flowing' counterpart performs significantly better than the static convolutional module.

Computer program product



[0112] In a third and fourth aspect, the invention concerns a computer program product comprising code instructions to execute a method (particularly on the data processing means 11a, 11b of the first or second server 1a, 1b) according to the first aspect of the invention for training a convolutional recurrent neural network, CRNN, or a method according to the second aspect of the invention for semantic segmentation in videos, and storage means readable by computer equipment (memory of the first or second server 1a, 1b) provided with this computer program product.


Claims

1. A method for training a convolutional recurrent neural network, CRNN, for semantic segmentation in videos;
the method being characterized in that it comprises the implementation, by a data processor (11a) of a first server (1a), of steps of:

(a) Training from a base of training images already semantically segmented, a first convolutional neural network, CNN;

(b) Training from a base of training videos already semantically segmented, a recurrent convolutional neural network, CRNN, corresponding to the first CNN wherein a convolutional layer has been replaced by a recurrent module having a hidden state; said training comprising, for each pair of successive frames (t - 1,t

1; T

2) of a video of said base of training videos already semantically segmented:

(b1) Warping the internal state of the recurrent layer according to an estimated optical flow between the frames of the pair, so as to adapt the internal state to the motion of pixels between the frames of the pair;

(b2) learning parameters of at least the recurrent module.


 
2. A method according to claim 1, wherein the pair of successive frames (t - 1, t

1; T

2) comprises a previous frame (t - 1) and a current frame (t), the optical flow estimated being the backward optical flow (Ft→t-1) from the current frame (t) to the previous frame (t - 1), sub step (a1) comprising applying a warping function to the internal state (ht-1) associated to the previous frame (t - 1) so as to obtain a warped internal state (ht-1) associated to the previous frame (t - 1) corresponding to the internal state (ht-1) wherein each pixel has underwent a displacement according to said backward optical flow (Ftt-1).
 
3. A method according to claim 2, wherein the warping function is such that ∀pt,ht-1(pt) = ht-1(pt + Ftt-1 (pt)), wherein pt is a pixel of the current frame t, Ftt-1 is the backward optical flow, ht-1 is the internal state associated to the previous frame t - 1, ht-1 is the warped internal state associated to the previous frame t - 1.
 
4. A method according to any one of claims 1 to 3, wherein step (b) comprises a previous sub-step (b0) of estimating the optical flow between the frames of the pair, performed using a second CNN, step (a) comprising training from a base of training pairs of successive frames between which the optical flow is known, the second CNN.
 
5. A method according to any claim 4, wherein said second CNN is a FlowNetSimple network or a FlowNetCorrelation network.
 
6. A method according to any one of claims 1 to 5, wherein the parameters of each layer of the CRNN before the recurrent module are fixed during sub-step (b2).
 
7. A method according to any one of claims 1 to 6, wherein the convolutional layer of the first CNN which is replaced by a recurrent module is the penultimate convolutional layer.
 
8. A method according to claims 6 and 7 in combination, wherein the parameters of the CRNN learnt at sub-step (b2) are those of the recurrent module and the last convolutional layer of the first CNN.
 
9. A method according to claims 4 and 8 in combination, wherein sub-step (b2) comprises further learning the parameters of the second CNN.
 
10. A method according to any one of claims 6 to 9, wherein the first CNN comprises before said penultimate convolutional layer an Atrous Spatial Pyramid Pooling module.
 
11. A method according to any one of claims 1 to 10, wherein the recurrent module is chosen among a convolutional Gated Recurrent Unit and a convolutional Long Short-Term Memory.
 
12. A method for semantic segmentation of an inputted video, characterized in that it comprises the implementation of steps of:

(a') Training, by data processing means (11a) of a first server (1a), from at least a base of training videos already semantically segmented, a recurrent convolutional neural network, CRNN, comprising a recurrent module having a hidden state;

(b') Performing, by processing means (11b) of a second server (1b), for each pair of successive frames (t - 1, t ∈

1; T

2) of the inputted video:

(b'0) estimating the optical flow between the frames of the pair;

(b'1) Warping the internal state of the recurrent layer according to the estimated optical flow, so as to adapt the internal state to the motion of pixels between the frames of the pair;

(b'2) Semantically segmenting each frame with the CRNN.


 
13. The method according to claim 12, wherein the step (a') is implemented in accordance with the learning method of any one of claims 1 to 11.
 
14. The method according to one of claims 12 and 13, wherein the step (a') comprises training from a base of training pairs of successive frames between which the optical flow is known, a convolutional neural network, CNN, sub-step (b'0) being performed using said CNN.
 
15. Computer program product comprising code instructions to execute a method according to one of claims 1 to 14 for training a recurrent convolutional neural network, CRNN, or for semantic segmentation of an inputted video, when said program is executed on a computer
 
16. A computer-readable medium, on which is stored a computer program product comprising code instructions for executing a method according to any one of claims 1 to 14 for training a recurrent convolutional neural network, CRNN, or for semantic segmentation of an inputted video.
 




Drawing
































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