The sparse facial shape descriptor extracted with traditional landmark-based face-tracker usually cannot capture the full details of the facial components’ shapes, which are essential to the recognition of higher level features such as facial expressions, emotions, identity, and so on. To overcome the limitations of sparse facial descriptors, we introduce the concept of face mask, a dense facial descriptor with information of semantic facial regions at pixel level like eyes and mouth. Developing from various deep learning-based semantic image segmentation methods, we then propose a novel approach for extracting face mask in video sequence. Different from semantic face segmentation, face mask extraction handles occlusion in a similar way to facial landmark tracking. Namely, the extract face mask is expected to be complete regardless of occlusion, while typical segmentation result would exclude the occluded area. Face mask extraction techniques could have many potential and interesting applications in the field of Human-Computer Interaction, including face detection & recognition, emotion & expression recognition, social robots interaction, etc. To the best of our knowledge, this is the first exploration of face mask extraction in video sequence with an end-to-end trainable deep-learning model.
Face mask extraction is a challenging task, especially for video clips taken in the wild, due to the huge amount of variations such as indoor & outdoor conditions, occlusions, image qualities, expressions, poses, skin colours, etc. Early studies of semantic face segmentation (Kae et al, 2013; Smith et al, 2013; Lee et al, 2008; Warrell and Prince, 2009) usually concentrated on the segmentation of still face images, and their methods were mostly based on heavily engineered approaches rather than learning.
In recent years, deep-learning techniques, particularly Convolutional Neural Networks (CNNs), has developed rapidly in the field of semantic image segmentation. Comparing to traditional engineering approaches, the major advantage of deep-learning methods is their ability to learn robust representations through an end-to-end trainable model for a particular task and dataset, and their performances usually surpass that of hand-crafted features extracted by traditional computer vision method. Among others, Fully Convolutional Networks (FCN)(Long et al, 2015) is the first seminal work of applying deep-learning techniques in semantic image segmentation. FCN substitute the fully connected layers in the widely-used deep CNN architectures - such as AlexNet (Krizhevsky et al, 2012), VGG-16 (Simonyan and Zisserman, 2014), GoogleLeNet (Szegedy et al, 2015), ResNet (He et al, 2016)
into convolutional layers, therefore turns the outputs from one-dimensional vectors to two-dimensional spatial heat-maps, which are then upsampled to the original image size using deconvolutional layers(Zeiler et al, 2011; Zeiler and Fergus, 2014). Developed from the baseline FCN, many improvements have been proposed in the following years, achieving increasingly better performance on benchmark datasets. Some works have changed the decoder structure of FCN, like SegNet by Badrinarayanan et al (2017), and some other models have applied Conditional Random Field (CRF) as a post-processing step, such as the CRFasRNN work by Zheng et al (2015) and the DeepLab models (Chen et al, 2016), and there are also works that utilise dilated convolutions (Zhou et al, 2015a), or atrous convolutions in other words, to broaden the reception fields of filters without additional computation cost, e.g. the DeepLab models by Chen et al (2016), ENet (Paszke et al, 2016) and the work of Yu and Koltun (2015).
Comparing to image segmentation, fewer works concern semantic segmentation in video sequences. Depending on the training methods, these works can be roughly divided into 1. fully-supervised methods (Kundu et al, 2016; Liu and He, 2015; Shelhamer et al, 2016; Tran et al, 2016; Tripathi et al, 2015), where all the annotations are given; 2. semi-Supervised approaches (Jain and Grauman, 2014; Nagaraja et al, 2015; Tsai et al, 2016; Caelles et al, 2017), which require certain pixel-level annotations like the ground truth of the sequence’s first frame; and 3. weakly-supervised ones (Saleh et al, 2017; Drayer and Brox, 2016; Liu et al, 2014; Wang et al, 2016), in which only the tags for each video clips are known. Due to the complex variations in real-life scenarios, we focus on fully-supervised video semantic segmentation. In addition, most semi-supervised or weakly-supervised approaches are proposed to solve the task of video object segmentation, i.e. binary classification between foreground and background, which limits their application in multi-class tasks such as face mask extraction.
To utilise the temporal information in video sequences, several fully-supervised video segmentation methods rely on graphical models such as Kundu et al (2016); Liu and He (2015); Tripathi et al (2015), while other approaches are based on CNN models, e.g. the Clockworks Convnets by Shelhamer et al (2016), in which a fixed or adaptive clock was used to control the update rates of different layers according to their semantic stability. Other works, such as Zhang et al (2014) and Tran et al (2016), use 3D convolutions or 3DCNNs to capture the temporal dependencies as well as the spatial connections. Both approaches have their limitations. Clockworks Convnets do not fully utilise the temporal information in video sequence since the semantic changes are only used to adjust clock rates. 3DCNN treats temporal dimension in the same way as 2D space, thus could limit the extraction of long-term temporal information.
In this paper, we propose an end-to-end trainable model which could exploit the temporal information in a more direct and natural way. The key idea is the application of Convolutional Long Short Term Memory (ConvLSTM) layer(Xingjian et al, 2015) in FCN models, which enable the FCNs to learn the temporal connections while retaining the ability to learn spatial correlations.
Recurrent Neural Networks, especially LSTMs, have already shown their capabilities to capture short and long term temporal dependencies in various computer vision tasks such as visual speech recognition (Lee et al, 2016; Zimmermann et al, 2016; Chung and Zisserman, 2016; Petridis et al, 2017b, a). However, typical RNN models only accept one-dimensional arrays, which limits the models’ application in tasks that require multi-dimensional relationships to be kept. To overcome this limitation, multiple approaches have been proposed, such as the works of Graves et al (2007), the ReNet architecture of Visin et al (2015), and the aforementioned ConvLSTM by Xingjian et al (2015).
Among these methods, ConvLSTM directly models the spatial relationships while keeping LSTM’s ability to capture temporal dependencies. Another advantage of ConvLSTM is it can be integrated into existing convolutional networks with very little effort because a convolutional layer can be easily replaced by a ConvLSTM layer with identical filter settings.
In this work, we introduce the ConvLSTM-FCN model that combines FCN and ConvLSTM by converting a certain convolutional layer in the FCN model into a ConvLSTM layer, thus adding the ability to model temporal dependencies within the input video sequence. Specifically, for the baseline model, we adopt the structure of FCN model based on ResNet-50 (He et al, 2016)
and then replace the classifying convolutional layer, which is converted from the fully connected layer in the original ResNet-50 model, with a ConvLSTM layer with the same convolutional filter settings. We also add two reshape layers since ConvLSTM layers require different input dimensions than the convolutional layers. The ConvLSTM-FCN model accepts video sequence as input and outputs the predictions of the same size, and the temporal information is learnt together with the spatial connections.
To be able to optimise the model toward higher accuracy in terms of mean Intersection over Union (mIoU), which is a typical performance metric for segmentation problems, we also propose a new loss function, called Segmentation Loss. Unlike the IoU loss in Rahman and Wang (2016), Segmentation Loss is more flexible and carries more practical meaning in image space. In comparison to the frequently-used cross-entropy loss, higher mIOU can be achieved when Segmentation Loss is used as the loss function during training.
A dataset with fully annotated face masks in videos would be needed to evaluate the proposed method. However, at this moment, no such dataset could be found in the public domain. Therefore, in this work, we use the 300 Videos in the Wild (300VW) dataset(Shen et al, 2015), which contains per-frame annotations of 68 facial landmarks for 114 short video clips. These landmark annotations are then converted into 4 semantic facial regions: face skin, eyes, outer mouth (lips) and inner mouth.
Our experiments are conducted on the aforementioned 300VW dataset with converted pixel-level labels of 5 class (the 4 facial regions plus background). As the baseline approaches, we compare performances of 1. The traditional 68-point facial landmark tracking model (Kazemi and Josephine, 2014); 2. The deeplab-V2 model (Chen et al, 2016); 3. The ResNet-50 Version FCN (He et al, 2016; Long et al, 2015), 4. The VGG-16 Version of FCN (Simonyan and Zisserman, 2014; Long et al, 2015). We then change the ResNet-50 version FCN to ConvLSTM-FCN, so that the temporal information in video sequence could be utilised. For better performance, we further extend our method to include three ConvLSTM-FCN models: a primary model to find the face region, and two additional models focusing on the eyes and mouth, respectively. The predictions of the three models are combined to obtain the final face mask. Our experimental results show that the utilisation of temporal information could significantly improve FCN’s performances for face mask extraction (from 54.50% to 63.76% mean IoU), and the performance of ConvLSTM-FCN model also surpass that of traditional landmark tracking models (63.76% Versus 60.09%).
2 Related Works
This section covers the major related works in the field. It is worth mentioning that, to the best of our knowledge, there is no similar work in terms of semantic face segmentation or face mask extraction in video sequence, so we have investigated the studies of video semantic segmentation instead.
2.1 Semantic Image Segmentation
The last few years have witnessed the rapid development of deep-learning techniques in the field of semantic image segmentation, and most of the state-of-the-art results are achieved by such models. The FCN by Long et al (2015) is the first milestone for deep learning in this field. FCN cast the fully convolutional layers in well-known deep architectures, such as AlexNet (Krizhevsky et al, 2012), VGG-16 (Simonyan and Zisserman, 2014), GoogleLeNet (Szegedy et al, 2015), ResNet (He et al, 2016), to convolutional layers so that the output of such models is spatial heat-maps instead of traditional one-dimensional class score. The skip-architecture of FCN enables the information from coarser layers to be seen by finer layers, therefore the model can be more aware of the global context, which is rather important in semantic segmentation. FCNs have limitations in term of integrating knowledge of the global context to make appropriate local predictions since the receptive field of their filters can only increase linearly when the number of layers grows (Garcia-Garcia et al, 2017). Therefore, later studies improve their models’ abilities to utilise the global image context with different approaches.
The works of the DeepLab models (Chen et al, 2016), ENet (Paszke et al, 2016) and the work of Yu and Koltun (2015) has involved the application of dilated convolutions, or so-called atrous convolutions. They are a kind of generalised Kronecker-factored convolutional filters (Zhou et al, 2015a), and they differ from traditional convolutional filters in that they have wider receptive fields which can grow exponentially with the dilated rate l (Garcia-Garcia et al, 2017). The standard convolutional operations can be seen as dilated convolutions with dilated rate = 1. Dilated convolutional layers can have more awareness of the global image context without reducing the resolution of feature maps too much. Another noticeable improvement is brought by the works of Yu and Koltun (2015), where their models take inputs of images at two different scales and then combine the predictions into one. The ideas of integrating predictions from multi-scale images can also be seen in the works of Roy and Todorovic (2016) and Bian et al (2016).
Conditional Random Field (CRF) is a frequently-used technique for deep semantic segmentation models, such as the DeepLab models (Chen et al, 2016) and the CRFasRNN by Zheng et al. Zheng et al (2015). The main advantage of CRF is that it could capture the long-range spatial relationships which are usually difficult for CNNs to retain, and CRF could also help to smooth the edges of the predictions.
2.2 Semantic Face Segmentation
Most earlier works of semantic face segmentation applied engineering-based approaches. Kae et al (2013)
employed a restricted Boltzmann machine to build the global-local dependencies such that the global shape can be natural, while they used CRFs to construct the details of the local shape. As in the work ofSmith et al (2013), a database of exemplary face images was first collected and labelled, and face images were aligned to those exemplary images with a non-rigid warping. There are also some other earlier works (Warrell and Prince, 2009; Scheffler and Odobez, 2011; Yacoob and Davis, 2006; Lee et al, 2008) in this field, however, most such works utilised engineering-based hand-crafted features, and it usually takes lots of time to fine-tune those models for them to work under particular scenarios. Therefore, they were gradually replaced by deep-learning based approaches.
Compared with the rapid progress of deep learning in semantic image segmentation, its application in semantic face segmentation is comparatively rare. Due to the difficulties of pixel-level labelling for huge amounts of data, currently, there are only a few publicly available datasets for this task. Two commonly used datasets are Parts Label dataset (Learned-Miller et al, 2016; Kae et al, 2013), which contains 2927 images with labels of background, face skin and hair, and Helen dataset (Le et al, 2012; Smith et al, 2013) including 2330 face images with annotations of face skin, left/right eyebrow, left/right eye, nose, upper lip, inner mouth, lower lip and hair. The lack of public face datasets with pixel-level annotations could be an obstacle for the development of deep models in this field.
For those face segmentation approaches using deep models, the works of Zhou et al (2015b) proposed an interlinked version of the traditional CNN model, where parts of the face could be detected except the facial skin. Compared with FCN, the proposed model is less efficient and its structure is overly redundant, and it cannot detect semantic part at large scales, like the facial skin. Güçlü et al (2017) took advantages of multiple deep-learning techniques, i.e. they formulated a CRF by one Convolutional Neural Network for the unary potential and the pairwise kernels, and one Recurrent Neural Networks to transform the unary potentials and the pairwise kernels into segmentation space. The training process utilised the idea of Generative Adversarial Networks (GAN), where the CRF and a discriminator network played a two-player minimised game. The limitation of this work is that it requires an initial face segmentation generated by a facial landmark detection model as the input in addition to the original face image, while the initial face segmentation is not necessary in our method.
All these semantic face segmentation approaches were proposed for still face images, while in the context of video sequences, where the variations are more complex, these methods may not be applicable. Currently, to the best of our knowledge, our work is the first one developed for semantic face segmentation in video sequence, or face mask extraction as we propose.
2.3 Video Semantic Segmentation
Video semantic segmentation methods can be roughly separated into three types through their supervision settings, which are: 1. The works that handle fully-supervised problems, i.e. the pixel-level annotations of all frames are known, 2. The semi-supervised video segmentation approaches, in which partial pixel-level annotations are known, such as only the ground-truths of the first frame is known for both training and testing, 3. The weakly-supervised methods focus on scenarios where only the tags of each video are given for the learning process. The main-stream interest of video segmentation community is on the semi-supervised problems (Jain and Grauman, 2014; Nagaraja et al, 2015; Tsai et al, 2016; Caelles et al, 2017) and the weakly-supervised issues (Saleh et al, 2017; Drayer and Brox, 2016; Liu et al, 2014; Wang et al, 2016), while the tasks of these problems are usually about segmenting one single object out of the background in a video sequence. This is somehow different from the scenarios of face mask extraction, where multiple semantic face parts should be extracted. Therefore, we have investigated the less-focused fully-supervised video segmentation works.
Some of these fully-supervised works replied on graphic models Kundu et al (2016); Liu and He (2015); Tripathi et al (2015). As for these approaches using deep models, the idea Clockworks Convnets by Shelhamer et al (2016) was based on the observation that the semantic contents of two successive frames change relatively slower than pixels. The proposed Clockworks Convnets used a clock at either fixed or adaptive schedules to control the update rates of different layers basing on the semantic content evolution. This work does not fully utilise the temporal information. The works of Zhang et al (2014) and Tran et al (2016) have both shown the idea of applying 3DCNN or 3D convolutions to capture information at time dimension. Treating temporal dependencies in the same way as spatial connections may hinder the model to understand some subtle temporal information, and they may not be able to capture the long-term time dependencies.
In our model, the temporal dependencies are extracted in a more natural and effective approach, through the application of Convolutional LSTM.
2.4 Convolutional LSTM
Convolutional LSTM (ConvLSTM) is proposed by Xingjian et al (2015) to solve the problem of precipitation nowcasting. Its has a similar structure as the FC-LSTM by Graves (2013), while all the inputs , …, , cell outputs , …, , hidden states , …, , input gate , forget gate and output gate
in ConvLSTM are 3D tensors, where the first dimension is the measurements in cell varying over time, and the last two dimension are spatial ones (rows and columns)(Xingjian et al, 2015). The key idea of ConvLSTM can be expressed in Eq. 1 (Xingjian et al, 2015), where ’’ denotes the convolutional operator and ’’ means the Hadamard product.
ConvLSTM could capture the long and short term temporal dependencies while retaining the spatial relationships in the feature maps, therefore it is an ideal candidate for face mask extraction in video sequence. Besides, with these convolutional operations in cells, a standard convolutional layer could be easily cast into a ConvLSTM layer with identical convolutional filters. Due to these advantages, we have utilised ConvLSTM in FCN structures to understand the temporal dependencies in video sequence.
The section explains our proposed ConvLSTM-FCN model and the segmentation loss function. In addition, we also introduce the engineering trick of combining the additional eye and mouth models with the primary model.
3.1 ConvLSTM-FCN Model
The first FCN model based on VGG-16 (Long et al, 2015) was proposed in 2015. Many variations of the FCN model have been developed afterward, usually achieving higher performances and better training efficiency.
In this work, we base our model on the structure of the FCN model released by Keras-Contributors (2018). This model is a ResNet-50 version FCN. The details about this model’s structure are summarised in Table 1. Compared with the standard ResNet-50 architecture (He et al, 2016), dilated convolutions with dilated rate = 2 are used in the building blocks of ’Conv5_x’ layer instead of the ordinary convolutional operations. The ’Conv6’ layer is the classifying layer which replaces the original fully-connected layer to produce feature maps of size 2020 at C channels, where C is the number of target classes. A bi-linear up-sampling layer of 16s is used instead of a deconvolutional layer.
|Layer Name||Building Blocks||Output Size||Dilated Rate|
7, 64, stride 2
3 max pooling, stride 2
|Conv6||11, C, stride 1||2020||11|
The conversion of baseline FCN to ConvLSTM-FCN is performed by replacing the ’Conv6’ layer with a ConvLSTM layer of identical convolutional filters. Fig. 1 shows the details of this procedure. The Reshape1 layer is used to output tensor with one additional time dimension ’T’, which is required by the ConvLSTM layer, and the Reshape2 layer cast the tensor back. ’T’, the time dimension in the ConvLSTM layer, refers to the number of frames in a video sequence. Therefore, for the ConvLSTM-FCN model to work effectively, the image orders within one batch should be arranged properly so that ConvLSTM layer could accept video sequences in the correct format.
3.2 Segmentation Loss
This section introduces the new loss function that we propose to optimise mean Intersection over Union (mIoU).
MIoU is the most frequently-used performance metric in the field of semantic segmentation. For one annotation set and its predictions, IoU is calculated by the intersection divided by the union. The intersection is actually the true positives of the confusion matrix, while the union is the sum of true positives, false positives and false negatives. mIoU is the average of IoUs over all classes. Assuming there are a total of C classes, and the notationstands for the number of pixels whose annotation is with prediction , then mIoU can be expressed in Eq. 2.
The main reason for using mIoU as the metric of segmentation accuracy instead of Classification Rate (CR) is to avoid the bias caused the class imbalances. Class imbalance is a common and challenging problem in semantic segmentation. For example, a face image usually contains much fewer eye pixels than background pixels. If all eye and background pixels are predicted as background, the resulting CR will still be quite high, which is unfair and misleading. In contrast, mIoU would be 0 in such case as there would be no true-positive for the eye pixels. Therefore, in the field of semantic segmentation, mIoU is used as the main evaluation metric, and its performance is not directly related to CR.
Cross-entropy loss, or softmax loss, is one of the most widely-used loss function in deep learning. Although cross-entropy loss is quite a powerful loss with smooth training curves, it targets toward higher average Classification Rate (CR), which does not necessarily lead to improvement in mIoU. Therefore, using cross-entropy loss in semantic segmentation could not fully fulfil deep models’ potential in the task. Therefore, we propose a new loss, which we name as Segmentation Loss, to optimise the model’s mIoU performances directly.
The work of Rahman and Wang (2016) has shown a similar idea of optimising IoU using an IoU loss instead of cross-entropy loss. This work, however, neglected the practical meaning of the IoU gradient, and, as a result, takes an over-simplified form. This will be shown in the following analysis.
Consider the case of single class segmentation, where annotations is either 1 (foreground, positive samples) or 0 (background, negative samples). Denote predictions as A, ground-truths as B and the network parameters as . Let and , then this single-class IoU can be expressed as in Eq. 3:
If we treat IoU as the direct objective function, we need to find IoU’s gradient, which is denoted as , in order to optimise this objective function. The deduction of is shown in Eq. 4.
The work of Rahman and Wang (2016) set the value of to 0 for pixels where ground-truths is 0, while is set to 0 for positive samples. However, we argue that the and , which is the gradient for and , hold their practical meanings in IoU optimisation and should not be simplified in this approach.
Since , for the purpose of optimising IoU, an appropriate gradient should encourage the predictions of the positive samples to change from 0 to 1. Similarly, for , the gradient () should make the values of negative samples’ predictions to vary from 1 to 0. From this perspective, stands for the optimisation direction of positive samples, while () reveals how to optimise negative samples. With these discoveries, we could reformulate the loss function regarding IoU in a meaningful and natural way. Assuming there are a total of samples and is the sample, if we let and , the proposed Segmentation Loss function can be found in Eq. 5.
In Eq. 5, and are the indicator functions for positive and negative samples respectively, and and are certain types of loss calculation functions for positive and negative samples separately.
Extending Eq. 5 to the case of total C classes and performing the normalisation, we can express the complete form of Segmentation Loss in Eq. 6. is now the indicator function for the positive samples of class , and vice versa for .
It can be seen from Eq. 6 that, in our Segmentation Loss, the loss of positive and negative samples from different classes is weighted separately by and , and these weights are somehow related to the number of samples over different classes. For example, if there are fewer samples belonging to class , its positive samples are more likely to hold a larger weight , since the union of class can be smaller than that of other classes. Therefore, our Segmentation Loss has properly considered the imbalanced data distributions over different classes, which are ignored in cross-entropy loss. Also, the Segmentation Loss is a more comprehensive loss definition for IoU optimisation when compared with the work of Rahman and Wang (2016).
The loss calculation function for positive and negative samples, which is and in Eq. 5 and Eq. 6, could have a variety of potential definitions. In this paper, we have provided two different definitions for them. Their first definition, which can be seen as a variant form of categorical hinge loss, is shown in Eq. 7.
In Eq. 7, and are both vectors, where is the model’s prediction for the sample , e.g. (-1.2,2.9,7.1) for a 3-class sample, and is the sample’s ground truth as a one-hot vector, such as (0,1,0) for a ground truth of 2 with total 3 classes. refers to the inverse of , for example, if , then . casts the number into the one-hot vector, and returns the maximum element. is a positive constant used to increase the discriminativities of loss function. The symbol ’’ represents vector’s Hadamard (element-wise) product, while ’’ means the dot product.
A second definition of and can be found in Eq. 8, where the meanings of , , and remain unchanged. The intuitions of this definition are straight-forward, encouraging the predicted values of ground truth class to increase and penalising for those false negative classifications.
3.3 Primary and Zoomed-in Models
In practice, to further increase segmentation accuracy, we have trained one primary model for initial face mask extraction and two additional models to focus on the eyes and mouth region, respectively. Particularly, the primary model takes a face video sequence and outputs face masks for each frame, and these face masks are used to localise and crop the eye and mouth regions out of the video sequence. Two additional trained models, one for eye and another for mouth region, are then used to generate the eye and mouth masks, which are usually more accurate than the corresponding regions in the primary face mask. The final predictions are obtained from the outputs of the three models, i.e. the eye and mouth masks are mapped back to the primary face mask, replacing these corresponding areas. The pipeline of how primary and additional models work is shown in Fig. 2.
All our experiments are implemented on the 300 Videos in the Wild (300VW) dataset (Shen et al, 2015). The 300VW dataset consists of 114 videos taken in unconstrained environments and the average duration of each video clip is 64 seconds with a frame rate of 30 fps. All 218595 frames in these videos have been annotated manually with the 68 facial landmarks as in the works of Sagonas et al (2013a, b). The scenarios of this dataset can be roughly divided into three categories with increasing challenges: 1. Category one where videos are taken under conditions with good lightings and potential occlusions such as glasses or beard may occur. 2. Videos of category two can have larger variations than category one, e.g. in-door environment without enough illumination, overly-exposed cameras, etc. while the occlusions are similar. 3. Category three is the most challenging one, with videos of high variations from totally unconstrained environments.
In order to obtain the face mask ground truths of all frames in the 300VW dataset, we have converted the 68-landmark annotations into pixel-level labels of one background class and four foreground classes: facial skin, eyes, outer mouth and inner mouth. This is achieved using cubic spline interpolation (with relaxed continuity constraints on eye corners and mouth corners) on corresponding landmark points. Some examples of the obtained face masks are shown in Fig.3. It can be seen from the figure that some videos in 300VW are quite challenging due to the high variations in head pose, illumination, occlusion, video resolution, etc.
After all the face masks have been generated, we have organised the dataset to suit our experiments. In particular, we have divided each video into short face sequences of one second (30 frames), and then for each video, we have randomly picked up 10% of its one-second sequences for our experiments. Since the information of adjacent one-second sequences may heavily overlap with each other, which may cause over-fitting problems, and also consider the training efficiency, we only use 10% one-second sequences instead of all these short clips. For training/validation/testing, we have randomly selected 619/58/80 one-second sequences, which contains 18570/1740/2400 face images in total, from 93/9/12 videos, and the training/validation/testing sets are subject-independent with each other to guarantee a fair evaluation. This dataset is called ’300VW-Mask’ dataset, and it is the dataset which we used to train the primary model and to evaluate the performance of final predictions.
For the training of these two additional models focusing on eye and mouth regions, we have further generated two sub-datasets from the afore-mentioned 300VW-Mask dataset. Specifically, we have cropped eye and mouth regions out of the 300VW-Mask dataset to form these sub-datasets. For the purpose of robustness, random noises are added during the cropping process, and we have fixed the locations of cropping box for every 5 consecutive frames so that the temporal information within these frames could be better extracted by the ConvLSTM-FCN models. Fig. 4 has plotted some examples of these two sub-datasets.
4.2 Experimental Framework
As mentioned in Section 3.2, mean Intersection over Union (mIoU) is used as the evaluation metric in the field of semantic segmentation, since mIoU is less sensitive to imbalanced data. Note that we ignored the IoU of background pixels in our mIoU calculation to focus the metric on the face mask pixels.
For the baseline approaches, we have compared the performances of the following methods on the 300VW-Mask dataset: 1. The traditional 68-point facial landmark tracking model (Kazemi and Josephine, 2014). 2. The deeplab-V2 model (Chen et al, 2016), 3. The ResNet-50 Version FCN (He et al, 2016; Long et al, 2015), and 4. The VGG-16 Version of FCN (Simonyan and Zisserman, 2014; Long et al, 2015).
For the facial landmark tracking model, we have used the 68-landmark model released by DLib library (King, 2009). This model has adopted the face alignment algorithm in the work of Kazemi and Josephine (2014), and have been trained on the iBUG 300-W face landmark dataset (Sagonas et al, 2016). We have implemented a 68-landmark face tracker with this alignment model using the methods described in Asthana et al (2014). This face tracker is run on all the testing set sequences, and the 68 output landmark points are then converted into face masks to calculate the mIoU performance, using the same conversion method as we used to generate face mask labels for the 300VW dataset.
Deeplab-V2 model is one of the most popular deep models in still image segmentation, and we have also evaluated the performance this model as one of the baseline methods. We have adopted the source code implementation released by Deeplab, and we have selected the model based on VGG-16 architecture.
The performances of FCN models are more relevant as our ConvLSTM-FCN model is based on the FCN architectures. Therefore, we have evaluated two different FCN models: 1. the ResNet-50 version FCN. This is the baseline FCN model that we adopted to convert into ConvLSTM-FCN. Section 3.1 described details about this FCN model and its conversion into ConvLSTM-FCN model. 2. the VGG-16 version FCN. This model has a similar overall architectures with the baseline FCN model, except that it is based on VGG-16.
Training ConvLSTM-FCN Models
Our ConvLSTM-FCN model, as mentioned in Section 3.1, is converted from the baseline FCN model by replacing the classification layer with ConvLSTM layers. Therefore, to simplify the training process, we first trained a baseline FCN model with all the training images without considering the temporal information. And then we converted this learned FCN model into ConvLSTM-FCN, keeping all the weights except the newly-added ConvLSTM layer, and then retrained it with data of video sequences, where the temporal correlations were learned and extracted.
In particular, the 300VW-Mask dataset was used to train the primary model. A baseline FCN was first trained on this dataset using cross-entropy loss, and this learned model was used as a reasonable starting point for the training of the primary ConvLSTM-FCN model. For the primary model, we have explored how the applications of ConvLSTM layer and Segmentation Loss could enhance the model’s performances by freezing all other layers except the ConvLSTM layer. After this exploration, we used Segmentation Loss to train the primary model by applying different learning rates on the ConvLSTM layer and other layers. Therefore, the training of the ConvLSTM-FCN model was performed as a two-step procedure: first, a baseline FCN model was trained with cross-entropy loss, then this learned model was converted to a ConvLSTM-FCN model to be trained with Segmentation Loss.
We have utilised similar training strategies for the additional eye and mouth models. Namely, we also first trained a baseline-FCN model focusing on the still eye and mouth images, and then a ConvLSTM-FCN with pre-trained weights was trained to capture the temporal dependencies.
We built and trained our model under the deep-learning frameworks of Keras (Chollet et al, 2015)
and TensorFlow(Abadi et al, 2015). The models are trained on a desktop with a 1080Ti graphics card and also on a cluster with 10 TITAN X graphics cards. It took around three days to obtain the final primary and additional models.
For the model training, we have adopted the Adam optimiser (Kingma and Ba, 2014)
, and model’s weights were saved and evaluated on the validation set after each epoch. The model with highest validation mIoU was then considered as the best one and was further evaluated on the testing set. All images were resized to 320 by 320 before they were fed into the model. For evaluations on the testing set, model’s output heat-map, whose size is also 320 by 320 pixels, was first resized back to the image’s original resolution, so that the IoU was calculated at this original scale.
The baseline model FCN was trained for a total of 80 epochs with batch size 16, learning rate 0.001 with linear decays and cross-entropy loss. The weights of the trained FCN model were then used as the starting point for the ConvLSTM-FCN model, which were trained for another 60 epochs using Segmentation Loss. The learning rate for ConvLSTM-FCN model was layer-based, which was 0.001 for the ConvLSTM layers and for other layers, where is a decaying factor for learning rate. The intuition is to train the newly-added ConvLSTM layer at larger steps while fine-tuning these learned layers with comparatively smaller learning rate.
For the ConvLSTM layer, the time dimension T was set to be 5, i.e. the ConvLSTM layer deals with short sequences of 5 frames. Therefore, input data of one batch should contain images, where N is an integer. In our experiments, we have set N=2, i.e. we have two 5-frame sequences in each batch.
In the step of integrating the predictions from primary and additional models, we first used the face masks from the primary model to approximately localise the eye and mouth regions for all frames, and then we fixed the cropping box of such regions for each 5-frame sequence so that the additional model could work smoothly to extract temporal information from these short sequences.
For each experiment, to verify its improvements on the baseline method, we also calculated whether it is statistically significant with the baseline FCN model. Particularly, we split the testing set, which contains 80 one-second sequences, into 10 groups, and calculated the P value of these 10 groups between the current experiment and the baseline model. If the P value is smaller than 0.05, then we consider this experimental result to be statistically significant from that of baseline approach.
Table 2 shows the performances of the four baseline approaches described in Section 4.2. The mIoU listed in the table is the average IoU of all classes except the background. It could be seen that although the face tracker approach has achieved the highest mIoU, its prediction for facial skin is worst than other deep methods. The performances of Deeplab-V2 model generally surpasses that of two FCN models, mainly on the eye and inner mouth predictions. These two FCN models achieved similar performances, giving the best facial skin predictions. All these deep models were trained with cross-entropy loss, and the trained model of FCN-ResNet50, which obtains 54.50% mIoU, would be converted into ConvLSTM-FCN model for further explorations. This trained model of FCN-ResNet50 will be simply called ’baseline-FCN’ for convenience.
Exploring ConvLSTM layer
As mentioned in Section 4.2, we have made some explorations in order to see if the ConvLSTM layer could actually improve the performance by using temporal information. For simplicity, after the baseline-FCN model was converted into ConvLSTM-FCN, we have frozen all other layers and only trained the newly-added ConvLSTM layer with cross-entropy loss. We have also tried two optimisers: Adam (Kingma and Ba, 2014) and RMSprop (Hinton et al, 2012). The results are shown in Table 3. It could be seen that the Adam and RMSprop optimisers both improve the mIoU slightly. For further validation, we have also computed their improvement over the baseline-FCN on the time dimension T, which is 5 in our ConvLSTM-FCN model. It could be seen in Table 4 that, for all 5-frame sequences, the improvements on the last four frames is generally higher than that of the first frame, which indicates the ConvLSTM layer can actually extract temporal information from video sequences to improve segmentation accuracy. Besides, it is also interesting to observe that the temporal smoothing effects are more obvious in the RMSprop experiment, with incremental improvements as time dimension increases.
Therefore, by these exploration experiments, we have verified that ConvLSTM could actually produce temporal smoothing effects for face mask extraction in video sequences. We have also selected Adam as the optimiser for following experiments.
|Definitions of ,||mIoU||FS||Eyes||OMT||IMT||BG|
|Eq. 7 (g=1)||58.10||90.90||51.96||59.32||30.20||94.82|
|Eq. 8 (g=0)||59.04||90.80||51.61||57.27||36.46||94.91|
|Eq. 8 (g=0.1)||58.39||90.56||51.45||57.43||34.11||94.76|
We have also conducted experiments to explore to what extend the proposed Segmentation Loss can lead to better a performance for the ConvLSTM-FCN model. As explained in Section 3.2, the loss calculation function for positive and negative samples, which is and in Eq. 6, could have various potential definitions, and we have provided two forms of them in Eq. 7 and Eq. 8. For the simplicity of the experiments, we have employed the same strategy as in the experiments of exploring ConvLSTM layer, i.e. after casting the baseline-FCN into ConvLSTM-FCN model, all other layers are frozen and the only trainable layer is the newly-added ConvLSTM layer. Then we used Segmentation Loss to train this partially-frozen ConvLSTM-FCN model. Table 5 summarises the results, and it could be seen that all results were improved (comparing to those shown in Table 3) when Segmentation Loss was used instead of cross-entropy loss. This demonstrates the effectiveness of the proposed Segmentation Loss in terms of optimising the ConvLSTM-FCN model. In addition, the loss function and defined Eq. 8 have shown the best mIoU performance when is 0, therefore, we have selected the form in Eq. 8 () for Segmentation Loss in the following experiments.
Training Primary and Zoomed-in Models
As mentioned in Section 4.2, We have applied similar strategies to train the primary and additional models. For the primary model, after the baseline-FCN was transformed into ConvLSTM-FCN, we have set different learning rates for different layers, which is 0.001 for ConvLSTM layer and for other layers, since we would like the newly-added ConvLSTM layer to learn faster than other already-trained layers. The Segmentation Loss with and defined in Eq. 8 (g=0) is used to train the primary ConvLSTM-FCN model. Table 6 has demonstrated the performances of the primary model with different values. It could be seen that different values could slightly affect the performances, while training ConvLSTM-FCN model with different internal learning rates could generally achieve better mIoUs than just freezing all layers except ConvLSTM layer.
Similarly, for the additional models on eye and mouth regions, we first used cross-entropy loss to train two baseline-FCN models on the eye and mouth sub-datasets, respectively, and these baseline-models are then converted into ConvLSTM-FCN models, which are also trained with different internal learning rates, as in the primary model’s training. Table 7 and Table 8 show the performances of baseline-FCN and ConvLSTM-FCN with different values. It can be seen from the results that ConvLSTM-FCN model with Segmentation Loss could generally improve the performance of the baseline-FCN model, and the additional model focusing on certain face region could achieve better segmentation accuracy on that region than that of the primary model.
|FCN-ResNet50 + cross-entropy||54.50||91.13||45.54||57.14||24.20||94.98|
|ConvLSTM-FCN (Freezing Other Layers) + cross-entropy||55.53||91.07||45.70||57.58||27.78||94.85|
|ConvLSTM-FCN (Freezing Other Layers) + Segmentation Loss||59.04||90.80||51.61||57.27||36.46||94.91|
|Primary Model + Segmentation Loss||60.04||90.46||54.89||58.98||35.86||94.36|
|Primary Model + Two Additional Models + Segmentation Loss||63.76||90.58||57.89||62.78||43.79||94.36|
As described in Section 3.3 and Section 4.2, the final predictions are obtained by integrating the face masks of the primary model, which provides localisations of eye and mouth regions, with the corresponding outputs of two additional models on the eye and mouth regions. These additional models focus on particular facial parts, such as eyes, outer and inner mouths, therefore they could produce more accurate segmentation results for these regions.
For the final predictions, we have used the primary model which are trained with , and the ConvLSTM models trained with for eye and mouth additional models (the performances of these models could be found in Table 6, Table 7 and Table 8).
The integration results could be found in Table 9, and this table also summarises the key improvements on the baseline-FCN model with different techniques. We can see from the table that combining primary model and additional models leads to a mIoU performance of 63.76%, which shows a 16.99% relative improvement on the baseline-FCN approach. Besides, when compared with these baseline approaches in Table 2, our proposed method still shows higher segmentation accuracies, even with the face tracker, which is the best-performing baseline approach.
In the task of face mask extraction, the temporal dimension carries important information which could be utilised to improve segmentation accuracies, especially when the information provided by current frame is not sufficient to allow reliable face mask extraction. This temporal-smoothing effect is what we would like to achieve with our ConvLSTM-FCN model.
In the case when normal FCN models encounter challenging segmentation tasks, the introduced ConvLSTM-FCN should be able to achieve better performances by exploiting information from both temporal and spatial domains. Fig. 5 plots some typical examples of such situations. As shown in the figure, the baseline-FCN model, which only learns the spatial relationships, have difficulties in segmenting face images with low qualities, occlusions, poor illuminations, etc. As a result, baseline-FCN could not effectively segment those smaller facial regions such as eyes and inner mouth under challenging scenarios. However, with the help of ConvLSTM-FCN model, the extracted face masks are more robust and realistic, especially for the smaller facial regions like eyes and inner mouth. The introduction of the zoomed-in model has further improved the segmentation results, which again verify the temporal-smoothing effects introduced by ConvLSTM-FCN.
Fig. 6 shows the mean IoU performances and standard deviation over all frames of each subject for the baseline-FCN, primary model and the integration of primary & additional models. The test set contains 80 one-second sequences coming from 12 videos, while these 12 videos are subject-independent with each other. It could be observed that the primary model or primary + additional have led to better performances than baseline-FCN on all the subjects. Besides, we could also see that the performances over different test subjects are generally similar, despite some fluctuations brought by the video variations.
In this paper, we have presented a novel ConvLSTM-FCN model for the task of face mask extraction in video sequences. We have illustrated how to convert a baseline-FCN model into ConvLSTM-FCN model, which can learn from both temporal and spatial domains. A new loss function named ’Segmentation Loss’ has also been proposed for training the ConvLSTM-FCN model. Last but not least, we also introduced the engineering trick of supplementing the primary model with two zoomed-in models focusing on eyes and moth. With all these are combined, we have successfully improved the performances of baseline-FCN on 300VW-Mask dataset from 54.50% to 63.76%, making a 16.99% relative improvement. The analysis of the experimental results has verified the temporal-smoothing effects brought by the ConvLSTM-FCN model.
Abadi et al (2015)
Abadi M, Agarwal A, Barham P, Brevdo E, Chen Z, Citro C, Corrado GS, Davis A, Dean J, Devin M, Ghemawat S, Goodfellow I, Harp A, Irving G, Isard M, Jia Y, Jozefowicz R, Kaiser L, Kudlur M, Levenberg J, Mané D, Monga R, Moore S, Murray D, Olah C, Schuster M, Shlens J, Steiner B, Sutskever I, Talwar K, Tucker P, Vanhoucke V, Vasudevan V, Viégas F, Vinyals O, Warden P, Wattenberg M, Wicke M, Yu Y, Zheng X (2015) TensorFlow: Large-scale machine learning on heterogeneous systems. URLhttps://www.tensorflow.org/, software available from tensorflow.org
Asthana et al (2014)
Asthana A, Zafeiriou S, Cheng S, Pantic M (2014) Incremental face alignment in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 1859–1866
- Badrinarayanan et al (2017) Badrinarayanan V, Kendall A, Cipolla R (2017) Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence 39(12):2481–2495
- Bian et al (2016) Bian X, Lim SN, Zhou N (2016) Multiscale fully convolutional network with application to industrial inspection. In: Applications of Computer Vision (WACV), 2016 IEEE Winter Conference on, IEEE, pp 1–8
- Caelles et al (2017) Caelles S, Maninis KK, Pont-Tuset J, Leal-Taixé L, Cremers D, Van Gool L (2017) One-shot video object segmentation. In: CVPR 2017, IEEE
- Chen et al (2016) Chen LC, Papandreou G, Kokkinos I, Murphy K, Yuille AL (2016) Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. arXiv preprint arXiv:160600915
- Chollet et al (2015) Chollet F, et al (2015) Keras. https://github.com/keras-team/keras
- Chung and Zisserman (2016) Chung JS, Zisserman A (2016) Out of time: automated lip sync in the wild. In: Asian Conference on Computer Vision, Springer, pp 251–263
- Drayer and Brox (2016) Drayer B, Brox T (2016) Object detection, tracking, and motion segmentation for object-level video segmentation. arXiv preprint arXiv:160803066
- Garcia-Garcia et al (2017) Garcia-Garcia A, Orts-Escolano S, Oprea S, Villena-Martinez V, Garcia-Rodriguez J (2017) A review on deep learning techniques applied to semantic segmentation. arXiv preprint arXiv:170406857
- Graves (2013) Graves A (2013) Generating sequences with recurrent neural networks. arXiv preprint arXiv:13080850
- Graves et al (2007) Graves A, Fernández S, Schmidhuber J (2007) Multi-dimensional recurrent neural networks. In: ICANN (1), pp 549–558
- Güçlü et al (2017) Güçlü U, Güçlütürk Y, Madadi M, Escalera S, Baró X, González J, van Lier R, van Gerven MA (2017) End-to-end semantic face segmentation with conditional random fields as convolutional, recurrent and adversarial networks. arXiv preprint arXiv:170303305
- He et al (2016) He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 770–778
- Hinton et al (2012) Hinton G, Srivastava N, Swersky K (2012) Rmsprop: Divide the gradient by a running average of its recent magnitude. Neural networks for machine learning, Coursera lecture 6e
- Jain and Grauman (2014) Jain SD, Grauman K (2014) Supervoxel-consistent foreground propagation in video. In: European Conference on Computer Vision, Springer, pp 656–671
- Kae et al (2013) Kae A, Sohn K, Lee H, Learned-Miller E (2013) Augmenting crfs with boltzmann machine shape priors for image labeling. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 2019–2026
- Kazemi and Josephine (2014) Kazemi V, Josephine S (2014) One millisecond face alignment with an ensemble of regression trees. In: 27th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, United States, 23 June 2014 through 28 June 2014, IEEE Computer Society, pp 1867–1874
- Keras-Contributors (2018) Keras-Contributors (2018) Keras-fcn website. URL https://github.com/aurora95/Keras-FCN, [Online; accessed 1-January-2018]
- King (2009) King DE (2009) Dlib-ml: A machine learning toolkit. Journal of Machine Learning Research 10:1755–1758
- Kingma and Ba (2014) Kingma DP, Ba J (2014) Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980
Krizhevsky et al (2012)
Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems, pp 1097–1105
- Kundu et al (2016) Kundu A, Vineet V, Koltun V (2016) Feature space optimization for semantic video segmentation. In: Computer Vision and Pattern Recognition (CVPR), 2016 IEEE Conference on, IEEE, pp 3168–3175
- Le et al (2012) Le V, Brandt J, Lin Z, Bourdev L, Huang TS (2012) Interactive facial feature localization. In: European Conference on Computer Vision, Springer, pp 679–692
- Learned-Miller et al (2016) Learned-Miller E, Huang GB, RoyChowdhury A, Li H, Hua G (2016) Labeled faces in the wild: A survey. In: Advances in face detection and facial image analysis, Springer, pp 189–248
- Lee et al (2016) Lee D, Lee J, Kim KE (2016) Multi-view automatic lip-reading using neural network. In: Asian Conference on Computer Vision, Springer, pp 290–302
- Lee et al (2008) Lee Kc, Anguelov D, Sumengen B, Gokturk SB (2008) Markov random field models for hair and face segmentation. In: Automatic Face & Gesture Recognition, 2008. FG’08. 8th IEEE International Conference on, IEEE, pp 1–6
- Liu and He (2015) Liu B, He X (2015) Multiclass semantic video segmentation with object-level active inference. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 4286–4294
- Liu et al (2014) Liu X, Tao D, Song M, Ruan Y, Chen C, Bu J (2014) Weakly supervised multiclass video segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 57–64
- Long et al (2015) Long J, Shelhamer E, Darrell T (2015) Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 3431–3440
- Nagaraja et al (2015) Nagaraja NS, Schmidt FR, Brox T (2015) Video segmentation with just a few strokes. In: ICCV, pp 3235–3243
- Paszke et al (2016) Paszke A, Chaurasia A, Kim S, Culurciello E (2016) Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:160602147
- Petridis et al (2017a) Petridis S, Wang Y, Li Z, Pantic M (2017a) End-to-end audiovisual fusion with lstms. arXiv preprint arXiv:170904343
- Petridis et al (2017b) Petridis S, Wang Y, Li Z, Pantic M (2017b) End-to-end multi-view lipreading. arXiv preprint arXiv:170900443
- Rahman and Wang (2016) Rahman MA, Wang Y (2016) Optimizing intersection-over-union in deep neural networks for image segmentation. In: International Symposium on Visual Computing, Springer, pp 234–244
- Roy and Todorovic (2016) Roy A, Todorovic S (2016) A multi-scale cnn for affordance segmentation in rgb images. In: European Conference on Computer Vision, Springer, pp 186–201
- Sagonas et al (2013a) Sagonas C, Tzimiropoulos G, Zafeiriou S, Pantic M (2013a) 300 faces in-the-wild challenge: The first facial landmark localization challenge. In: Computer Vision Workshops (ICCVW), 2013 IEEE International Conference on, IEEE, pp 397–403
- Sagonas et al (2013b) Sagonas C, Tzimiropoulos G, Zafeiriou S, Pantic M (2013b) A semi-automatic methodology for facial landmark annotation. In: Computer Vision and Pattern Recognition Workshops (CVPRW), 2013 IEEE Conference on, IEEE, pp 896–903
- Sagonas et al (2016) Sagonas C, Antonakos E, Tzimiropoulos G, Zafeiriou S, Pantic M (2016) 300 faces in-the-wild challenge: Database and results. Image and Vision Computing 47:3–18
- Saleh et al (2017) Saleh FS, Aliakbarian MS, Salzmann M, Petersson L, Alvarez JM (2017) Bringing background into the foreground: Making all classes equal in weakly-supervised video semantic segmentation. In: 2017 IEEE International Conference on Computer Vision (ICCV), IEEE, pp 2125–2135
- Scheffler and Odobez (2011) Scheffler C, Odobez JM (2011) Joint adaptive colour modelling and skin, hair and clothing segmentation using coherent probabilistic index maps. In: British Machine Vision Association-British Machine Vision Conference, EPFL-CONF-192633
- Shelhamer et al (2016) Shelhamer E, Rakelly K, Hoffman J, Darrell T (2016) Clockwork convnets for video semantic segmentation. In: Computer Vision–ECCV 2016 Workshops, Springer, pp 852–868
- Shen et al (2015) Shen J, Zafeiriou S, Chrysos GG, Kossaifi J, Tzimiropoulos G, Pantic M (2015) The first facial landmark tracking in-the-wild challenge: Benchmark and results. In: Proceedings of the IEEE International Conference on Computer Vision Workshops, pp 50–58
- Simonyan and Zisserman (2014) Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:14091556
- Smith et al (2013) Smith BM, Zhang L, Brandt J, Lin Z, Yang J (2013) Exemplar-based face parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 3484–3491
- Szegedy et al (2015) Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, Rabinovich A (2015) Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 1–9
- Tran et al (2016) Tran D, Bourdev L, Fergus R, Torresani L, Paluri M (2016) Deep end2end voxel2voxel prediction. In: Computer Vision and Pattern Recognition Workshops (CVPRW), 2016 IEEE Conference on, IEEE, pp 402–409
- Tripathi et al (2015) Tripathi S, Belongie S, Hwang Y, Nguyen T (2015) Semantic video segmentation: Exploring inference efficiency. In: SoC Design Conference (ISOCC), 2015 International, IEEE, pp 157–158
- Tsai et al (2016) Tsai YH, Yang MH, Black MJ (2016) Video segmentation via object flow. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp 3899–3908
- Visin et al (2015) Visin F, Kastner K, Cho K, Matteucci M, Courville A, Bengio Y (2015) Renet: A recurrent neural network based alternative to convolutional networks. arXiv preprint arXiv:150500393
- Wang et al (2016) Wang H, Raiko T, Lensu L, Wang T, Karhunen J (2016) Semi-supervised domain adaptation for weakly labeled semantic video object segmentation. In: Asian Conference on Computer Vision, Springer, pp 163–179
- Warrell and Prince (2009) Warrell J, Prince SJ (2009) Labelfaces: Parsing facial features by multiclass labeling with an epitome prior. In: Image Processing (ICIP), 2009 16th IEEE International Conference on, IEEE, pp 2481–2484
- Xingjian et al (2015) Xingjian S, Chen Z, Wang H, Yeung DY, Wong WK, Woo Wc (2015) Convolutional lstm network: A machine learning approach for precipitation nowcasting. In: Advances in neural information processing systems, pp 802–810
- Yacoob and Davis (2006) Yacoob Y, Davis LS (2006) Detection and analysis of hair. IEEE transactions on pattern analysis and machine intelligence 28(7):1164–1169
- Yu and Koltun (2015) Yu F, Koltun V (2015) Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:151107122
- Zeiler and Fergus (2014) Zeiler MD, Fergus R (2014) Visualizing and understanding convolutional networks. In: European conference on computer vision, Springer, pp 818–833
- Zeiler et al (2011) Zeiler MD, Taylor GW, Fergus R (2011) Adaptive deconvolutional networks for mid and high level feature learning. In: Computer Vision (ICCV), 2011 IEEE International Conference on, IEEE, pp 2018–2025
- Zhang et al (2014) Zhang H, Jiang K, Zhang Y, Li Q, Xia C, Chen X (2014) Discriminative feature learning for video semantic segmentation. In: Virtual Reality and Visualization (ICVRV), 2014 International Conference on, IEEE, pp 321–326
- Zheng et al (2015) Zheng S, Jayasumana S, Romera-Paredes B, Vineet V, Su Z, Du D, Huang C, Torr PH (2015) Conditional random fields as recurrent neural networks. In: Proceedings of the IEEE International Conference on Computer Vision, pp 1529–1537
- Zhou et al (2015a) Zhou S, Wu JN, Wu Y, Zhou X (2015a) Exploiting local structures with the kronecker layer in convolutional networks. arXiv preprint arXiv:151209194
- Zhou et al (2015b) Zhou Y, Hu X, Zhang B (2015b) Interlinked convolutional neural networks for face parsing. In: International Symposium on Neural Networks, Springer, pp 222–231
- Zimmermann et al (2016) Zimmermann M, Ghazi MM, Ekenel HK, Thiran JP (2016) Visual speech recognition using pca networks and lstms in a tandem gmm-hmm system. In: Asian Conference on Computer Vision, Springer, pp 264–276