This paper proposes a discriminative feature learning approach for deep face recognition using a center loss function in addition to softmax loss. The center loss aims to learn discriminative features that reduce intra-class variations. It works by minimizing distances between feature vectors and their corresponding class centers, which are updated during training. Experimental results on benchmarks like LFW, YTF, and MegaFace demonstrate state-of-the-art performance for face verification and identification tasks when using the proposed softmax loss combined with center loss. While performance improvements are achieved, the paper also acknowledges there is still room for enhancing results to meet practical demands involving large-scale datasets with millions of distractors.

1. A Discriminative Feature Learning Approach
for Deep Face Recognition
Yandong Wen, Kaipeng Zhang, Zhifeng Li and Yu Qiao
Paper Seminar @ SK Telecom: Jisung Kim

2. Separable vs Discriminative

3. A toy example : What’s wrong with Softmax

4. A toy example : What’s wrong with Softmax
1. m = mini-batch size
2. n = the number of class
3. x_i = feature vector in R^d (d is the feature dimension)
4. W = R^(dxn), b = R^n (bias)

5. Is it good for clustering?
Separable,
the deep features are not discriminative enough.
by intra-class variation

6. Is it good for clustering?
Training Set (50K) Test Set (10K)

7. Let’s be discriminative by Center Loss
1. m = mini-batch size
2. c_yi = yth class center in d dimension
3. x_i = feature vector in R^d (d is the feature dimension)
4. But! c_yi should be updated as the deep features changed.
5. Average the features of every class in each iteration.

8. Let’s be discriminative by Center Loss
1. Two modification
a. Updating the centers based on mini-batch
b. Updating center with learning rate alpha (for mislabelled samples)
2. Total = Softmax Loss + lambda * Center Loss

9. Varying λ with Loss = Soft + λ*Center

10. The discriminative feature learning algorithm

11. CNN Architecture → https://github.com/ydwen/caffe-face

12. Compared to Siamese and Triplet
1. Dramatic data expansion
a. Make (x_i, x_j) pair or (x_positive, x_negative, x_anchor) pair
2. Hard to make proper pair → Hard to make decreasing loss → Hard to train

13. The Devil is in the details.
1. Preprocessing
2. Training Data
3. Detailed Settings in CNNs
4. Detailed Settings in Testings

14. Implementation Details : Preprocessing
1. Cropped Face Size = (112x96x3)
a. Subtraction 127.5
b. Dividing by 128
c. Every Pixel Values in -1 <= val <= 1
2. Face Detect by MTCNN
3. Use 5 landmarks
a. Two Eyes
b. One Nose
c. Two Mouth Corners
4. Alignment : Similarity Transformation.
a. Rotation, Translation, Scaling

15. Implementation Details : Training data
1. 17,189 unique persons
2. 0.7 M images
a. CASIA-WebFace : 0.49 M, 10,575
b. CACD2000 : 0.16 M, 2,000
c. Celebrity+ : 0.20 M, 10,177
3. Removing Same Identities !!!
4. Data Augmentation : Only horizontally Flipped

16. Implementation Details : Detailed Settings in CNNs
1. Model A : Softmax Loss
2. Model B : Softmax Loss & Contrastive Loss
3. Model C : Softmax Loss & Center Loss
4. Batch Size : 256
5. GPUs : 2 x Titan X
6. Learning rate : Start from 0.1
7. Model A, C
a. Divided by 10 at the 16K, 24K
b. Complete at 28K (Roughly costs 14 hours)
8. Model B
a. Divided by 10 at the 24K, 36K
b. Complete at 42K (Roughly costs 22 hours)

17. Implementation Details : Detailed Settings in Testing
1. Deep Features : first FC layer
2. Extract 2 feature
a. Original Image
b. Horizontally Flipped Image
3. Just Concatenate
4. Do PCA
5. Score by cosine distance
6. Identification
a. Nearest Neighbor
7. Verification
a. Threshold Comparison

18. λ & α on LFW

19. LFW (images) & YTF (videos)

20. Verification Performance (λ=0.003 & α=0.5)

21. MegaFace

22. MegaFace
1. Gallery Set : 690,000 persons / 1 milion Images ( Distractors )
2. Probe set
a. Facescrub : 530 persons / 100K images
b. FGNet : 82 persons / 1,002 images (ages varying from 0 to 69)
3. Small / Large Protocol
a. When training db size < 0.5M, < 20K persons
b. When training db size > 0.5M, > 20K persons
4. Face Identification
a. From 1 vs 10
b. To 1 vs 1,000,000
5. Face Verification
a. 4 bilion = 4,000,000,000 negative pairs

23. MegaFace : Identification (1M distractors)

24. MegaFace : Verification (10^-6 FAR, 1M distractors)

25. FaceScrup
1. 530 celebrities
2. 100k images

26. FGNet
1. 82 persons
2. 1002 images

27. MegaFace : Identification (FaceScrub)

28. MegaFace : Verification (FaceScrub)

29. MegaFace : Identification (FGNet)

30. MegaFace : Verification (FGNet)

31. Conclusions
1. USE!!! Center Loss for getting discriminative feature
2. Much room for performance improvement.
a. To meet practical demand
b. Identification : Rank-1 with 1M distractor
c. Verification : 10^-6 FAR with 1M distractor