Research map / Label-efficient and robust learning / Recognition and long-tail
SphereFace2: Binary Classification is All You Need for Deep Face Recognition
Read the paperCheck 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 |
|---|---|---|---|
| Add positive/negative sample balance (ablation) | SphereFace2 binary classification · Combined validation set | verification accuracy: 74.49 -> 78.03 | Helped |
| Add easy/hard sample mining (ablation) | SphereFace2 with positive/negative balance · Combined validation set | verification accuracy: 78.03 -> 89.65 | Helped |
| Add angular margin (ablation) | SphereFace2 with balance and easy/hard mining · Combined validation set | verification accuracy: 89.65 -> 93.87 | Helped |
| Add similarity adjustment (ablation) | SphereFace2 with balance, easy/hard mining, and angular margin · Combined validation set | verification accuracy: 93.87 -> 94.28 | Helped |
| Set λ=0.8 rather than λ=0.3 (ablation) | SphereFace2 with t=1 · Combined validation set | verification accuracy: 92.72 -> 93.87 | Helped |
| Enable similarity adjustment (t=3 rather than t=1) (ablation) | SphereFace2 · Combined validation set | verification accuracy: 93.73 -> 94.28 | Helped |
| Increase angular margin from m=0.2 to m=0.3 (ablation) | SphereFace2 with λ=0.7, r=30, t=3 · Combined validation set | verification accuracy: 93.93 -> 94.52 | Helped |
| Use similarity adjustment with CosFace (ablation) | CosFace · IJB-B 1:1 verification, FAR=1e-5 | TAR: 79.35 -> 75.72 | Hurt |
| Use SphereFace2 instead of the strongest listed baseline for each metric | Face recognition models · IJB-B 1:1 verification, FAR=1e-5 | TAR: 80.75 -> 85.40 | Helped |
| Train with pair-wise labels rather than class-wise labels | SphereFace2 · Combined validation set with noisy training labels | verification accuracy | Helped |
| Replace CosFace-type margin with ArcFace-type or multiplicative margin (ablation) | SphereFace2 · Combined validation set | verification accuracy: 94.28; 94.19; 94.19 | No clear effect |
| Apply similarity adjustment to multi-class loss (ablation) | CosFace and ArcFace · IJB-B and IJB-C | TAR and TPIR: CosFace 79.35 -> 75.72; ArcFace 78.12 -> 78.88 | Mixed |
Closest papers by meaning
- SphereFace: Deep Hypersphere Embedding for Face Recognition (CVPR 2017 · similarity 0.89)
- Softmax Dissection: Towards Understanding Intra- and Inter-Class Objective for Embedding Learning (AAAI 2020 · similarity 0.87)
- ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · similarity 0.87)
- GIF: Generative Inspiration for Face Recognition at Scale (CVPR 2025 · similarity 0.85)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.84)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.84)
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.