Face Ensemble

未分类

Leaderboard of LFW (labeled face in wild)

- Metric Learning: 10000 individuals

• Feature Extraction
  
  • Lightened-CNN
  • Softmax-Face
  • Center-Face

• Test-Cos-LFW (feature 512)

  • Lightened-CNN

  • Softmax-Face

  • Center-Face

    | Model         | PCA Size | Threshold | Score  |
    | ------------- | -------- | --------- | ------ |
    | Lightened CNN | 160      | 0.63      | 98.78% |
    | Softmax Face  | 240      | 0.61      | 99.25% |
    | Center Face   | 192      | 0.65      | 99.43% |
    

• Concatenation + PCA (feature 1024)

  • Softmax-Face + Center-Face + Lighten-CNN

  • Softmax-Face + Center-Face

  • Center-Face + Lightened-CNN

    | Model             | PCA Size | Threshold | Score  |
    | ----------------- | -------- | --------- | ------ |
    | Soft+Cent+Lighten | 320      | 0.62      | 99.38% |
    | Softmax+Center    | 384      | 0.62      | 99.37% |
    | Center+Lighten    | 192      | 0.65      | 99.43% |
    | Softmax+Lighten   | 384      | 0.61      | 99.27% |
    

• Class Balanced

  • Only Softmax

  • Center-Softmax

    | Model            | PCA Size | Threshold | Score  |
    | ---------------- | -------- | --------- | ------ |
    | Balanced Center  | 144      | 0.65      | 99.33% |
    | Balanced Softmax | 192      | 0.62      | 99.08% |
    

• Single Model Improvement by Learning methods

  ​:x: Not working: Test 92.4% :arrow_right: 93.5% LFW 99.43% :arrow_right: 99.33%

  ​:x: The Objectiveness Dissimilarity inside.

• Metric Learning on mild subset

  • Clean CISIA-WebFace: 10,575 subjects and 494,414 images with 27,703 deleted. (More eastern faces)
    - [ ] detect subject with less than 10 faces: 10458 subjects with 455594 faces
  • Clean MS-data 1W: Sample subset 10,000 subjects with most face images
• Test on Real data

• Mirror Face
• More Ensembles
• Patch learning
• Improve Lightened CNN by center-loss
• NormFace (happynear)
• More structures