Video face recognition has become more and more significant in the past few years [43, 41, 33, 21, 28, 29, 44, 30, 22, 26, 40, 12, 46, 11, 34], which plays an important role in many practical applications such as visual surveillance, access control, person identification, video search and so on. Compared to single still image-based face recognition, further useful information of a single face can be exploited in the video. However, the video faces exhibit much richer uncontrolled variations, , out-of-focus blur, motion blur, occlusion, varied illuminations and a large range of pose variations, which make video face recognition a challenging task. Hence, how to design a feature model which can effectively represent the video face across different frames becomes a key issue of video face recognition.
In video face recognition task, each subject usually has a varied number of face images. A straightforward approach would be to represent a video face as a set of face descriptors extracted by a deep neural network, compare every pair of face descriptors between two face videos [33, 36], and fuse the matching results across all pairs. However, this method would be considerably memory-consuming and inefficient especially for a large-scale recognition task. Consequently, an effective aggregation scheme, requiring minimal memory storage and supporting efficient similarity computation, is desired for this task, to generate a compact representation for a face video. And what is more, the aggregated representations should be discriminative, , they are expected to have smaller intra-class distance than inter-class distance under a suitably chosen metric space.
. Besides max pooling, average pooling [21, 28, 34, 8] may be the most common aggregation technique. However, it considers all frames of equal importance during feature aggregation, in which case the low quality frames with some misleading feature would degrade the performance of recognition. Considering of this problem, some other methods either just focus on high quality frames, , feature-rich frames, while ignoring low quality frames, such as blurred faces, occluded faces and large pose faces [29, 12] or adaptively high weigh high quality frames while framesdown weigh low quality frames [44, 46].
Despite that those aggregation strategies have been shown to be effective in the previous works, we believe that an optimal aggregation strategy should not simply and crudely despise the low quality frames, because the low quality frames might even contain local discriminative features which can be complementary to high quality frames. In some sense, the low quality frames may be beneficial to video face recognition. Thus, the best aggregation result should be the composition of local discriminative features from low quality frames and other parts from high quality frames. Our intuition is simple and straightforward: an ideal algorithm should be able to emphasize the valuable part of the frame feature while suppress the worthless part of the frame feature irrespective of the face quality during aggregation, , it adaptively deals with each dimension of frame feature with different importance, not like NAN  that treats each dimension of equal importance for frame feature when aggregating. Let us imagine an extreme case: with some poor quality face images, , a variety of large pose faces each with different pose, it is possible to aggregate these faces into a discriminative face representation for video face recognition.
To this end, we propose a new attention-based aggregation network which adaptively and fine-grained weighs the feature along each feature dimension among all frames to form a compact and discriminative face representation. Different from previous methods, we neither focus only on high quality frames nor simply weigh the feature on frame-level. Instead, we design a neural network which is able to adaptively and fine-grained measure the importance of each dimension of the feature among all frames.
Our major contributions can be summarized as follows:
We propose a novel feature aggregation scheme for video face recognition, and reveal why it could work better. It is a generalized feature aggregation scheme, and may also serve as a feature aggregation scheme for other computer vision tasks.
Based on the proposed aggregation scheme, we construct a feature aggregation network (as shown in Figure 1) composed of two modules trained end-to-end or one by one separately. One is the feature embedding module which is a frame-level feature extractor using deep CNN model. The other is the aggregation module which adaptively integrates the feature vectors of all the video frames together. Our feature aggregation network inherits the main advantages of the pooling techniques (, average pooling and max pooling), could handle arbitrary input size and produces order-invariant, fixed-size feature representation.
We demonstrate the effectiveness of our proposed aggregation scheme in video face recognition by various comparative experiments. Trained on publicly available dataset, such as YouTube face dataset and IJB-A dataset, our method takes a lead over the baseline methods and is a competitive method compared to the state of the art methods.
2 Related works and preliminaries
Since our work is concerned with order-insensitive video or image set face recognition, any other methods exploiting the temporal information of video sequence will not be considered here.
Early traditional studies attempt to represent the face videos or image sets as manifolds [1, 17, 19, 38, 37, 39] or convex hulls  and compute their similarities under corresponding spaces. While those methods may work well under constrained scenarios, they are usually incapable of dealing with large face variations.
Some other methods extract the local features of frames and aggregate them across multiple frames to represent the videos [21, 20, 27]. For example, PEP-based methods [21, 20] take a part-based representation by extracting and merging LBP or SIFT descriptors, and the method in  applies Fisher vector encoding to represent each frame by extracting RootSIFT [3, 25] and fuses across multiple different video frames to form a video-level representation.
These years, still image-based face recognition has gained great success thank to deep learning techniques[33, 40, 36, 10, 23]. Based on this, some simple aggregation strategies are adopted in video face recognition. The methods in  and  utilize pairwise frame feature similarity computation and then fuse the matching results. Max- or average-pooling is used to aggregate the frame features in [28, 11, 7, 8]. Though DAN  proposes a GAN-like aggregation network which takes the video clip as input and reconstructs a single image as output to represent the video, the average pooling result of the video frames is employed to supervise the aggregation training. What is more, DAN is not suitable to tackle image set face recognition due to that a video face discriminator is used inside the GAN.
Recently, a few methods take a lead over the simple pooling techniques. The method in  utilizes discrete wavelet transform and entropy computation to select feature-rich frames from a video sequence and learns a joint feature from them. GhostVLAD  employs a modified NetVLAD  layer to down weigh the contribution of low quality frames. NAN  proposes an attention mechanism to adaptively weigh the frames, so that the contribution of low quality frames to the aggregation is down weighed. However, NAN considers each dimension of the feature vector to be of equal importance. These methods may lose some valuable information of the low quality images. This motivates us to seek a better solution in this paper.
Our work is inspired by NAN . However, our aggregation scheme is a more generalized strategy, can fine-grained handle the feature vector on dimension level. Now, let us review the feature aggregation scheme of NAN . Consider the video face recognition task on pairs of video face data , where is a face video sequence or image set with varying image number , , in which is the -th frame in the video, and is the corresponding subject ID of . Each frame has a corresponding normalized feature representation extracted from the feature embedding module, and the aggregated feature representation becomes
where is the linear weight generated from all feature vectors of a video, it can be formulated as
where is the corresponding significance yielded via dot product with a kernel filter for each feature vector, it can be formulated as
Obviously, if , Eq. (1) degrades to average pooling strategy.
3.1 The proposed aggregation scheme
We argue that each dimension of the feature vector shares the common weight as NAN does is not optimal. The ideal strategy should be able to adaptively weigh each dimension of feature vector separately. So we leverage a kernel matrix to filter the feature vector via product, yielding a significance vector , which describes the importance of each dimension of . Assuming is an -dimension vector, then, we can formulate as
and formulate as
After softmax operation along each dimension, a positive weight vector is generated as following
where denotes the linear weight of that -th dimension of the feature vector contributes to aggregation result, and . So that the aggregated feature representation becomes
where represents element-wise product. Figure 2 shows the calculating process of . turns out to be after -normalization. Either cosine or distance can be used to compute the similarity.
From above formulas and Figure 2, we can clearly see the difference between our method and NAN is that we use a kernel matrix instead of a kernel vector to adaptively weigh the feature. Therefore, we can measure the importance of feature on dimension level without constraining each dimension to share the same weight just as NAN  does. Compared to NAN and other pooling techniques, our method is more flexible, and can make each dimension of one feature vector adaptively contribute to the aggregation feature. In theory, it can realize optimal feature aggregation after well trained. So, our method can deal with every frame fairly regardless of face quality, and make the best to exploit its any valuable or discriminative local feature to promote the video face recognition.
3.2 The proposed feature aggregation network
Based the on proposed aggregation scheme, we construct a feature aggregation network comprised of two modules. As shown in Figure 1, the network can be fed with a set of face images of a subject and produces a single feature vector as its representation for the recognition task. It is built upon a modern deep CNN model for frame feature embedding, and adaptively aggregates all frames in the video into a compact vector representation.
The image embedding module of our network adopts the backbone network of Arc-Face  which greatly advances the image-based face recognition recently. The embedding module mainly consists of a ResNet50 which has an improved residual unit: BN-Conv-BN-PReLu-Conv-BN structure, while using BN-Dropout-FC-BN after the last convolutional layer. The embedding module produces 512-dimension image features which are first normalized to be unit vectors then fed into the aggregation module.
In order to obtain a better aggregation representation, a cascaded two attention blocks with nonlinear transfer is designed inside aggregation module as shown in Figure 3. Each attention block consists of a kernel filter and a nonlinear transfer. The kernel filter is implemented with a layer, while nonlinear transfer with a activation layer. Then becomes
where is the output of the first block, it can be formulated as
Therefore, besides kernel matrices and , biases and are also trainable parameters of aggregation module. We have to point out that our cascaded attention blocks are totally different from NAN ’s in that our attention block uses an importance matrix while NAN uses an importance vector to weigh the feature vectors. In comparison, our method is more fine-grained than NAN to aggregate feature vectors. Furthermore, NAN aggregates the feature vectors twice, where the second attention block takes the aggregation result of the first attention block as input. However, our method only makes aggregation once.
In addition, our network has several other favorable properties. First, it is able to tackle arbitrary number of images for one subject. Second, the aggregation result which is of the same size as a single feature
is invariant to the image order, keeps unchanged when the image sequence are reshuffled or even reversed, , our network is insensitive to the temporal information of the video or image set. Third, it is adaptive to the input faces and whose all parameters are trainable through supervised learning with standard backpropagation and gradient descent.
3.3 Network training
To make the training faster and more stable, we divide it into three stages(as shown in Fig.4). Firstly, we train the embedding module for single image face recognition task. In this stage, the cleaned MS-Celeb-1M dataset [10, 13] is used. Secondly, we train the whole network end-to-end for set-based face recognition task, and the VGGFace2 dataset  is used in this stage. In order to boost the capability of handling images of varying quality that typically occur in the wild, the VGGFace2 datatset is augmented in the form of image degradation, such as blurring or compression. Finally, we finetune the whole network end-to-end on the training set of the benchmark dataset.
4.1 Datasets and protocols
4.2 Training details
Embedding module training: As aforementioned, the cleaned MS-Celeb-1M dataset [10, 13] which contains about 3.8M images of 85k unique identities is used to train our feature embedding network for the single image face recognition task. MTCNN  is employed to detect 5 facial landmarks in the face images. The faces are aligned to by using similarity transformation according to the landmarks detected, and then fed into embedding network for training. The Additive Angular Margin Loss , which is a kind of modified softmax loss is used to supervise the training. After training, the classification loss layer is removed from the trained network. The rest network is fixed and used to extract a single fixed-size representation for the face image.
End-to-End training: We use the VGGFace2 dataset  to train the whole network end-to-end for the set-based face recognition task. The VGGFace2 Dataset  consists of about 3 million images, covering 8631 identities, and there are on average 360 face images for each identity. To perform set-based face recognition training, the image sets are built by repeatedly sampling a fixed number of images which belong to the same identity. All the images sampled are aligned using the same way as in the embedding module training. After alignment, the data augmentation is performed by image degradation. Following the same strategy as in GhostVLAD ,four methods: isotropic blur, motion blur, decreased resolution and JPEG compression are adopted to degrade the face image for training. The Additive Angular Margin Loss is also adopted to supervise the end-to-end training. In order to speed up the training, we initialize all the parameters of the aggregation module to be zero. That means the aggregation module begins with average pooling to search the optimal parameters.
Finetuning: All the video face dataset are also aligned by using MTCNN  algorithm and similarity transformation.Then the whole network is finetuned on the training set of each video face dataset using the Additive Angular Margin Loss .
|Wen et al.||94.90||92.30|
4.3 Baseline methods
Since average pooling is a widely used aggregation method in many previous works [21, 28, 34, 8], we choose average pooling as one of our baselines. For fairness, the average pooling method shares common embedding module with our method after the whole network is finetuned on each benchmark dataset. We also choose NAN  as our another baseline. We reproduce the NAN which consists of cascaded two attention blocks as  describes. The reproduced NAN is trained in the same way as our method. The two baselines as well as our method produce 512-d feature representation for each video and compute the similarity in time. Besides the above two baselines, we also compare with some other sate-of-the-art methods.
|Method||1:1 verification TAR(%)|
|DREAM||86.8 1.5||94.4 0.9||-|
|Triplet Embedding||81.3 2||91 1||96.4 0.5|
|Template Adaptation||83.6 2.7||93.9 1.3||97.9 0.4|
|NAN||88.1 1.1||94.1 0.8||97.8 0.3|
|QAN||89.31 3.92||94.20 1.53||98.02 0.55|
|VGGFace2||92.1 1.4||96.8 0.6||99.0 0.2|
|GhostVLAD||93.5 1.5||97.2 0.3||99.0 0.2|
|AvgPool||88.82 1.22||96.18 0.92||98.16 0.40|
|NAN*||93.12 1.16||96.91 0.83||98.71 0.599|
|Ours||93.61 1.51||97.28 0.28||98.94 0.31|
|Method||1:N identification TPIR(%)|
|DREAM||-||-||94.6 1.1||96.8 1.0||-|
|Triplet Embedding||75.3 3||86.3 1.4||93.2 1||-||97.7 0.5|
|Template Adaptation||77.4 4.9||88.2 1.6||92.8 1.0||97.7 0.4||98.6 0.3|
|NAN||81.7 4.1||91.9 0.9||95.8 0.5||98.0 0.5||98.6 0.3|
|VGGFace2||88.3 3.8||94.6 0.4||98.2 0.4||99.3 0.2||99.4 0.1|
|GhostVLAD||88.4 5.9||95.1 0.5||97.7 0.4||99.1 0.3||99.4 0.2|
|AvgPool||86.43 4.81||94.05 1.02||95.69 0.62||98.52 0.45||99.04 0.33|
|NAN*||87.92 5.44||94.83 1.01||97.23 0.57||99.05 0.58||99.24 0.44|
|Ours||88.51 5.86||95.18 1.02||97.92 0.32||99.23 0.36||99.39 0.25|
4.4 Results on YouTube Face Dataset
We first evaluate our method on the YouTube Face Dataset  which contains 3425 videos of 2595 different subjects. The lengths of videos vary from 48 to 6070 frames and the average length is 181.3 frames per video. The dataset is splitted into 10 folds, and each fold consists of 250 positive (intra-subject) pairs and 250 negative (inter-subject) pairs. We follow the standard verification protocol to test our method.
Table 1 shows the results of our method, the baseline and some other state of the art methods. The ROC curves of our method and the baselines are shown in Figure 5. We can see that our method outperforms the two baselines, reducing the error of the best-performing baseline:NAN* by 6.88%. Our method also performs better than all the other state-of-the-art methods (including the original NAN  )except the deep FR methods and ADRL method . The reason is that, the deep FR method benefits a lot from front face selection and triplet loss embedding with carefully selected triplets, and ADRL method  benefits from exploiting the temporal information from the video sequence. Compared to deep FR method, our aggregation method is more straightforward and elegant without hand-crafted rules. And compared to ADRL method , our method is order-invariant, can be used in more potential scenarios. It is noteworthy that our reproduced NAN also performs better than original NAN . That is because both the embedding module and aggregation module of the reproduced NAN is trained end-to-end instead of separately, and compared to separate training, more training data is used during the end-to-end training stage.
4.5 Results on IJB-A Dataset
The IJB-A Dataset  contains 5712 images and 2085 videos, covering 500 subjects in total. The average numbers of images and videos per subject are 11.4 images and 4.2 videos. This dataset is more challenging than YouTube Face dataset  due to it covers large range of pose variations and different kinds of image conditions.
We follow the standard benchmark procedure for IJB-A to evaluate our method on both the ‘compare’ protocol for face verification and the ‘search’ protocol for face identification. The true accept rates (TAR) vs. false positive rates (FAR) are reported for verification, while the true positive identification rates (TPIR) vs. false positive identification rates (FPIR) and the Rank-N accuracies are reported for identification. Table 2 and Table 3 show the evaluation results of different methods for verification task and identification task respectively. And Figure 6 shows the receiver operating characteristics (ROC) curves for verification as well as the cumulative match characteristic (CMC) and decision error trade-off (DET) curves for identification.
From Table 2, we can see that our method outperforms the two baselines, reducing the error of best-performing baseline by 7.12%,11.97% and 17.83% at FAR=0.001, FAR=0.01 and FAR=0.1 respectively in verification task. Compared to all the state-of-the-art methods, our method performs a little better at FAR=0.001 and FAR=0.01, and performs on par with them at FAR=0.01 where TAR values have almost saturated to a 99% mark. And From Table 3, we also can see that our method performs better than the two baselines by appreciable margins, and beats all the state of the art methods except on Rank-10 metric where our method is on par with them and the TPIR values have saturated to a 99.4% mark. Besides, the reproduced NAN also outperforms the original NAN  as on YTF benchmark. It is noteworthy that the gap between our method and the original NAN  on IJB-A dataset is larger compared to the results on YTF dataset. This is because the face variations in IJB-A dataset is much larger than in YTF dataset, our method can extract more beneficial information for video face recognition.
We introduced a new feature aggregation network for video face recognition. Our network can adaptively and fine-grained weigh the input frames along each dimension of the feature vector and fuse them organically into a compact representation which is invariant to the frame order. Our aggregation scheme can make the best to exploit any valuable part of the features regardless of the frame quality to promote the performance of video face recognition. Experiments on YTF and IJB-A benchmarks show our method is a competitive aggregation method.
O. Arandjelovic, G. Shakhnarovich, J. Fisher, R. Cipolla, and T. Darrell.
Face recognition with image sets using manifold density divergence.
Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on, volume 1, pages 581–588. IEEE, 2005.
-  R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5297–5307, 2016.
-  R. Arandjelovic and A. Zisserman. Three things everyone should know to improve object retrieval. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2911–2918. IEEE, 2012.
-  K. Cao, Y. Rong, C. Li, X. Tang, and C. Change Loy. Pose-robust face recognition via deep residual equivariant mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5187–5196, 2018.
-  Q. Cao, L. Shen, W. Xie, O. M. Parkhi, and A. Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018), pages 67–74. IEEE, 2018.
-  H. Cevikalp and B. Triggs. Face recognition based on image sets. In Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on, pages 2567–2573. IEEE, 2010.
-  J.-C. Chen, R. Ranjan, A. Kumar, C.-H. Chen, V. M. Patel, and R. Chellappa. An end-to-end system for unconstrained face verification with deep convolutional neural networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 118–126, 2015.
-  A. R. Chowdhury, T.-Y. Lin, S. Maji, and E. Learned-Miller. One-to-many face recognition with bilinear cnns. In 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1–9. IEEE, 2016.
-  N. Crosswhite, J. Byrne, C. Stauffer, O. Parkhi, Q. Cao, and A. Zisserman. Template adaptation for face verification and identification. Image and Vision Computing, 79:35–48, 2018.
-  J. Deng, J. Guo, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. arXiv preprint arXiv:1801.07698, 2018.
-  C. Ding and D. Tao. Trunk-branch ensemble convolutional neural networks for video-based face recognition. IEEE transactions on pattern analysis and machine intelligence, 40(4):1002–1014, 2018.
-  G. Goswami, M. Vatsa, and R. Singh. Face verification via learned representation on feature-rich video frames. IEEE Transactions on Information Forensics and Security, 12(7):1686–1698, 2017.
-  Y. Guo, L. Zhang, Y. Hu, X. He, and J. Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In European Conference on Computer Vision, pages 87–102. Springer, 2016.
-  T. Hassner, I. Masi, J. Kim, J. Choi, S. Harel, P. Natarajan, and G. Medioni. Pooling faces: Template based face recognition with pooled face images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 59–67, 2016.
-  J. Hu, J. Lu, and Y.-P. Tan. Discriminative deep metric learning for face verification in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1875–1882, 2014.
-  J. Hu, J. Lu, J. Yuan, and Y.-P. Tan. Large margin multi-metric learning for face and kinship verification in the wild. In Asian Conference on Computer Vision, pages 252–267. Springer, 2014.
-  T.-K. Kim, O. Arandjelović, and R. Cipolla. Boosted manifold principal angles for image set-based recognition. Pattern Recognition, 40(9):2475–2484, 2007.
-  B. F. Klare, B. Klein, E. Taborsky, A. Blanton, J. Cheney, K. Allen, P. Grother, A. Mah, and A. K. Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1931–1939, 2015.
-  K.-C. Lee, J. Ho, M.-H. Yang, and D. Kriegman. Video-based face recognition using probabilistic appearance manifolds. In Computer vision and pattern recognition, 2003. proceedings. 2003 ieee computer society conference on, volume 1, pages I–I. IEEE, 2003.
-  H. Li, G. Hua, Z. Lin, J. Brandt, and J. Yang. Probabilistic elastic matching for pose variant face verification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3499–3506, 2013.
-  H. Li, G. Hua, X. Shen, Z. Lin, and J. Brandt. Eigen-pep for video face recognition. In 2014 Asian Conference on Computer Vision (ACCV), pages 17–33, 2014.
-  L. Liu, L. Zhang, H. Liu, and S. Yan. Toward large-population face identification in unconstrained videos. IEEE Transactions on Circuits and Systems for Video Technology, 24(11):1874–1884, 2014.
-  W. Liu, Y. Wen, Z. Yu, M. Li, B. Raj, and L. Song. Sphereface: Deep hypersphere embedding for face recognition. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, page 1, 2017.
-  Y. Liu, J. Yan, and W. Ouyang. Quality aware network for set to set recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5790–5799, 2017.
-  D. G. Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91–110, 2004.
-  O. M. Parkhi, K. Simonyan, A. Vedaldi, and A. Zisserman. A compact and discriminative face track descriptor. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1693–1700, 2014.
-  O. M. Parkhi, K. Simonyan, A. Vedaldi, and A. Zisserman. A compact and discriminative face track descriptor. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1693–1700, 2014.
-  O. M. Parkhi, A. Vedaldi, A. Zisserman, et al. Deep face recognition. In BMVC, volume 1, page 6, 2015.
-  Y. Rao, J. Lin, J. Lu, and J. Zhou. Learning discriminative aggregation network for video-based face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3781–3790, 2017.
Y. Rao, J. Lu, and J. Zhou.
Attention-aware deep reinforcement learning for video face recognition.In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3931–3940, 2017.
-  Y. Rao, J. Lu, and J. Zhou. Attention-aware deep reinforcement learning for video face recognition. In Proceedings of the IEEE International Conference on Computer Vision, pages 3931–3940, 2017.
-  S. Sankaranarayanan, A. Alavi, C. D. Castillo, and R. Chellappa. Triplet probabilistic embedding for face verification and clustering. In 2016 IEEE 8th international conference on biometrics theory, applications and systems (BTAS), pages 1–8. IEEE, 2016.
-  F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 815–823, June 2015.
-  K. Sohn, S. Liu, G. Zhong, X. Yu, M.-H. Yang, and M. Chandraker. Unsupervised domain adaptation for face recognition in unlabeled videos. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3210–3218, 2017.
-  Y. Sun, X. Wang, and X. Tang. Deeply learned face representations are sparse, selective, and robust. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2892–2900, 2015.
-  Y. Taigman, M. Yang, M. Ranzato, and L. Wolf. Deepface: Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1701–1708, 2014.
-  P. Turaga, A. Veeraraghavan, A. Srivastava, and R. Chellappa. Statistical computations on grassmann and stiefel manifolds for image and video-based recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(11):2273–2286, 2011.
-  R. Wang, S. Shan, X. Chen, and W. Gao. Manifold-manifold distance with application to face recognition based on image set. In Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on, pages 1–8. IEEE, 2008.
W. Wang, R. Wang, Z. Huang, S. Shan, and X. Chen.
Discriminant analysis on riemannian manifold of gaussian distributions for face recognition with image sets.In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2048–2057, 2015.
-  Y. Wen, K. Zhang, Z. Li, and Y. Qiao. A discriminative feature learning approach for deep face recognition. In European Conference on Computer Vision, pages 499–515. Springer, 2016.
-  L. Wolf, T. Hassner, and I. Maoz. Face recognition in unconstrained videos with matched background similarity. In 2011 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 529–534, June 2011.
-  L. Wolf, T. Hassner, and I. Maoz. Face recognition in unconstrained videos with matched background similarity. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 529–534. IEEE, 2011.
-  L. Wolf and N. Levy. The svm-minus similarity score for video face recognition. In 2013 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3523–3530, 2013.
-  J. Yang, P. Ren, D. Zhang, D. Chen, F. Wen, H. Li, and G. Hua. Neural aggregation network for video face recognition. In CVPR, volume 4, page 7, 2017.
-  K. Zhang, Z. Zhang, Z. Li, and Y. Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters, 23(10):1499–1503, 2016.
-  Y. Zhong, R. Arandjelović, and A. Zisserman. Ghostvlad for set-based face recognition. arXiv preprint arXiv:1810.09951, 2018.