Research map / Architectures and efficiency / RNNs and normalization
An Efficient Training Approach for Very Large Scale Face Recognition
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 FC layer with DCP | face recognition training on MS1MV2, Glint360k, and Webface42M · 9-benchmark average | average: 97.29 -> 96.80; 97.67 -> 96.83; 98.15 -> 97.90 | Hurt |
| combine identity-based and instance-based loaders (ablation) | F²C with MobileFaceNet · MS1MV2, IJBC | TPR@FAR=1e-4: 82.30 -> 91.06; 89.30 -> 91.06 | Helped |
| combine identity-based and instance-based loaders (ablation) | F²C with MobileFaceNet · MS1MV2, MegaFace | FPR@FAR=1e-6: 79.19 -> 89.30; 86.40 -> 89.30 | Helped |
| use twin Probe and Gallery nets rather than a single net (ablation) | F²C with MobileFaceNet · MS1MV2, Small Datasets | average accuracy: 93.90 -> 95.29 | Helped |
| use twin Probe and Gallery nets rather than a single net (ablation) | F²C with MobileFaceNet · MS1MV2, MegaFace | rank-1 identification: 82.69 -> 89.30 | Helped |
| increase DCP features per identity from K=1 to K=2 (ablation) | F²C with MobileFaceNet · MS1MV2, Small Datasets | average accuracy: 95.19 -> 95.29 | Helped |
| increase DCP features per identity from K=1 to K=2 (ablation) | F²C with MobileFaceNet · MS1MV2, MegaFace | FPR@FAR=1e-6: 88.31 -> 89.30 | Helped |
| use the default 1:1 instance-to-identity loader ratio (ablation) | F²C with ResNet50 · MS1MV2, MegaFace | FPR@FAR=1e-6: 92.74 -> 96.74; 93.65 -> 96.74; 96.43 -> 96.74; 90.56 -> 96.74 | Helped |
| use only the identity-based loader rather than dual loaders (ablation) | F²C with MobileFaceNet · MS1MV2, Small Datasets | average accuracy: 95.29 -> 94.20 | Hurt |
| use only the instance-based loader rather than dual loaders (ablation) | F²C with MobileFaceNet · MS1MV2, Small Datasets | average accuracy: 95.29 -> 94.24 | Hurt |
Closest papers by meaning
- Dynamic Class Queue for Large Scale Face Recognition in the Wild (CVPR 2021 · similarity 0.90)
- Virtual Fully-Connected Layer: Training a Large-Scale Face Recognition Dataset With Limited Computational Resources (CVPR 2021 · similarity 0.89)
- Accelerated Training for Massive Classification via Dynamic Class Selection (AAAI 2018 · similarity 0.88)
- Killing Two Birds With One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC (CVPR 2022 · similarity 0.87)
- WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition (CVPR 2021 · similarity 0.86)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.85)
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.