Research map / Label-efficient and robust learning / Recognition and long-tail
Killing Two Birds With One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the full FC class-center set with positive centers and a random negative subset at ratio 0.3 (ablation) | ResNet50 face recognition training on WebFace42M · MFR-All | TAR@FAR=1e-6 (%): 93.86 -> 94.03 | Helped |
| Use partial negative-center sampling at ratio 0.1 instead of full FC | ResNet100 training on WebFace42M · WebFace42M, 2M identities, 8 GPUs | throughput (samples/second): 2463 -> 3552 | Helped |
| Use partial negative-center sampling at ratio 0.1 instead of full FC | ResNet100 training on WebFace42M · WebFace42M, 2M identities, 8 GPUs | GPU memory (GB): 18.9 -> 11.8 | Helped |
| Reduce the sampling ratio to 0.04 (ablation) | PFC on WebFace4M · MFR-All | TAR@FAR=1e-6 (%): 86.25 -> 74.11 | Hurt |
| Reduce the sampling ratio to 0.3 (ablation) | PFC on WebFace4M · MFR-All | TAR@FAR=1e-6 (%): 86.25 -> 86.85 | Helped |
| Reduce the sampling ratio to 0.013 (ablation) | PFC on WebFace12M · MFR-All | TAR@FAR=1e-6 (%): 91.70 -> 87.85 | Hurt |
| Reduce the sampling ratio to 0.3 (ablation) | PFC on WebFace12M · MFR-All | TAR@FAR=1e-6 (%): 91.70 -> 91.82 | Helped |
| Reduce the sampling ratio to 0.008 (ablation) | PFC on WebFace42M · MFR-All | TAR@FAR=1e-6 (%): 93.86 -> 91.27 | Hurt |
| Reduce the sampling ratio to 0.3 (ablation) | PFC on WebFace42M · MFR-All | TAR@FAR=1e-6 (%): 93.86 -> 94.03 | Helped |
| Use partial negative-center sampling at ratio 0.1 rather than full FC (ablation) | ResNet50 trained on WebFace12M-Conflict · MFR-All | TAR@FAR=1e-6 (%): 79.93 -> 91.20 | Helped |
| Add abnormal inter-class filtering at cosine similarity 0.4 to PFC-0.2 (ablation) | ResNet50 trained on WebFace12M-Conflict · MFR-All | TAR@FAR=1e-6 (%): 90.55 -> 91.68 | Helped |
| Use partial negative-center sampling at ratio 0.1 rather than full FC (ablation) | ResNet50 trained on WebFace12M-Flip(40%) · MFR-All | TAR@FAR=1e-6 (%): 43.87 -> 78.53 | Helped |
Closest papers by meaning
- Virtual Fully-Connected Layer: Training a Large-Scale Face Recognition Dataset With Limited Computational Resources (CVPR 2021 · similarity 0.90)
- Mis-Classified Vector Guided Softmax Loss for Face Recognition (AAAI 2020 · similarity 0.90)
- CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018 · similarity 0.89)
- Softmax Dissection: Towards Understanding Intra- and Inter-Class Objective for Embedding Learning (AAAI 2020 · similarity 0.88)
- More Information Supervised Probabilistic Deep Face Embedding Learning (ICML 2020 · similarity 0.87)
- An Efficient Training Approach for Very Large Scale Face Recognition (CVPR 2022 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.