![Train Data : MSCeleb-1M (10M images for 100K identities / refined version (3.8M images for
85k identities)
Test Data : nine commonly used datasets (LFW, YTF, CALFW, CPLFW, CFP, AgeDB-30,
MegaFace, IJB-B and IJB-C)
Metrics : TAR@FAR = 1e-6 / TAR@FAR=1e-5 (Compare a True Accept Rate at a certain
False Accept Rate)
Baseline Model : ArcFace(ResNet-100) GDN(two blocks of MLP[BN-FC] and a FC for group
classification)
Learning : 8 synchronized GPUs and a mini-batch involving 128 images per GPU
learning rate : 0.005 for the first 50k, 0.0005 for the 20k, and 0.00005 for 10k with a weight
decay of 0.0005 and a momentum of 0.9 with stochastic gradient descent (SGD)
Pre-trained network : only the softmax-based loss(Arcface).(To stable the group-probability)
Experiments](https://image.slidesharecdn.com/papergroupface-210124134318/85/GroupFace-Face-Recognition-13-320.jpg)
GroupFace (Face Recognition)
- 1. Susang Kim(healess1@gmail.com) Face Recognition GroupFace: Learning Latent Groups and Constructing Group-based Representations for Face Recognition
- 2. Face Recognition(Recap) Deep Learning boosts the performance. sparse-representation-based classifier DEEP FR SYSTEM WITH FACE DETECTOR AND ALIGNMENT. 1) A FACE DETECTOR IS USED TO LOCALIZE FACES. 2) THE FACES ARE ALIGNED TO NORMALIZED CANONICAL COORDINATES. 3) FACE ANTI-SPOOFING RECOGNIZES WHETHER THE FACE IS LIVE OR SPOOFED; 4) FACE PROCESSING IS USED TO HANDLE VARIATIONS BEFORE TRAINING AND TESTING, (E.G. POSES, AGES; DIFFERENT ARCHITECTURES AND LOSS FUNCTIONS TO EXTRACT DISCRIMINATIVE DEEP FEATURE WHEN TRAINING) FACE MATCHING METHODS ARE USED TO DO FEATURE CLASSIFICATION AFTER THE DEEP FEATURES OF TESTING DATA ARE EXTRACTED. Deep Face Recognition: A Survey: https://arxiv.org/pdf/1804.06655.pdf
- 3. Deep Face Recognition: A Survey : https://arxiv.org/pdf/1804.06655.pdf Face recognition focused on improving the loss function. 1) Softmax Loss and its Variations (Deepface) 2) Euclidean-distance-based Loss (FaceNet) 3) Angular/cosine-margin-based Loss (Sphereface, arcface) ⇒ +GroupFace(group-aware representation + Self-distributed Face Recognition research is highly saturated
- 4. Racial Faces in-the-Wild (RFW) database proved that existing commercial APIs and the SOTA algorithms indeed work unequally for different races and the maximum difference in error rate between the best and worst groups is 12%, as shown in Table VIII. Racial Faces in-the-Wild : https://arxiv.org/pdf/1812.00194.pdf Validated the racial bias (RFW dataset - ICCV 2019) If the faces used to train the algorithm are predominantly white men, the system will have a harder time recognizing anyone who doesn't fit.
- 5. Kakao에서 쓴 얼굴인식 최신 논문 (CVPR 2020)으로 종래에 Metric Learning을 통한 Loss개선 중 심에서 Group Learning의 추가 적용을 통해 각 Group 간의 sample 수를 Uniform하게 분 포하게 함으로써 Search Space를 줄일 수 있게 되었고 1:1 face verification과 1:N face identification tasks에서 Public Data로 LFW, YTF, CALFW, CPLFW, CFP, AgeDB- 30, MegaFace, IJB-B and IJB-C 부문 SOTA 기록 https://arxiv.org/pdf/2005.10497.pdf Face Recognition Paper (CVPR 2020)
- 6. 유사한 얼굴의 Group Learning을 통해 Identity 별 Search Space 적용 Summary of Group Learning
- 7. Each person has own characteristics in his or her face. At the same time, they have common ones shared in a group of people. In the real world, group-based description (man with deep, black eyes and red beard) that involves common characteristics in the group, can be useful to narrow down the set of candidates, even though it cannot identify the exact person. We propose a novel face-recognition architecture called GroupFace that learns multiple latent groups and constructs group-aware representations to effectively adopt the concept of grouping Instance + Group = Enriched Feature 종래방식
- 9. ResNet-100 5 points normalized face (112x112) Follow Arcface/Cosface to set the hyper- parameters of the loss function(softmax- based) identity label given sample embedding feature function number of identities Instance-based Representation
- 10. Accuracy Practical BN+FC Group-aware Representation Structure group-aware similarity cosine similarity (instance) distance metric (group) The parameters are determined empirically to be β = 0.1 and γ = 1/3. Group-aware Similarity
- 11. Naive Labeling (non-uniform group) Self-distributed Labeling Self-distributed Grouping normalized probability K groups Expected value = 0 (to generate uniformly-distributed group labels)
- 12. arcface loss self-grouping loss empirically set to 0.1 SphereFace : https://arxiv.org/pdf/1704.08063.pdf Comparison of decision boundaries in binary case. Note that, θi is the angle between Wi and x. Loss Function for Learning
- 13. Train Data : MSCeleb-1M (10M images for 100K identities / refined version (3.8M images for 85k identities) Test Data : nine commonly used datasets (LFW, YTF, CALFW, CPLFW, CFP, AgeDB-30, MegaFace, IJB-B and IJB-C) Metrics : TAR@FAR = 1e-6 / TAR@FAR=1e-5 (Compare a True Accept Rate at a certain False Accept Rate) Baseline Model : ArcFace(ResNet-100) GDN(two blocks of MLP[BN-FC] and a FC for group classification) Learning : 8 synchronized GPUs and a mini-batch involving 128 images per GPU learning rate : 0.005 for the first 50k, 0.0005 for the 20k, and 0.00005 for 10k with a weight decay of 0.0005 and a momentum of 0.9 with stochastic gradient descent (SGD) Pre-trained network : only the softmax-based loss(Arcface).(To stable the group-probability) Experiments
- 14. GroupFace is also effective for a lightweight model such as ResNet-34 that requires only 8.9 GFLOPS less than ResNet-100, which requires 24.2 GFLOPS. ResNet-34 based GroupFace shows a similar performance of ResNet-100 based ArcFace and greatly outperforms ResNet-100 in a most difficult criterion (FAR=1e-6). Ablation Studies
- 15. Self-distributed grouping ⇒ uniform distribution Group-aware representations ⇒ search space GDN : self group + latent group ⇒ 어느정도 불균형은 있으나 전반적으로 균일함 Evaluations
- 16. Example images belonging to each groups. As enormous identities (80k∼) of large scale dataset cannot be mapped to a few groups (32), each group contains identities of multiple characteristics. Some groups have one common visual description (Group 5: Some Men, Group 20: Bald Men) with some variations while others have multi-mode visual descriptions. GroupFace is summarized in two main ways: (1) It is well known that additional supervisions from different objectives can bring an improvement of the given task by sharing a network for feature extraction, a segmentation head can improve accuracy in object detection. Likewise, learning the groups can be a helpful cue to train a more generalized feature extractor for face recognition. (2) GroupFace proposes a novel structure that fuses instance based representation and group-based representation, which is empirically proved its effectiveness. Conclusion
- 17. Thanks Any Questions? You can send mail to Susang Kim(healess1@gmail.com)