Research map / Label-efficient and robust learning / Recognition and long-tail
ArcFace: Additive Angular Margin Loss for Deep 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 standard softmax target logit with a target logit using an additive angular margin | Face recognition with CASIA and ResNet50 · CFP-FP | verification accuracy (%): 94.39 -> 95.56 | Helped |
| Use ArcFace margin 0.5 rather than margins 0.4, 0.45, or 0.55 (ablation) | ArcFace on CASIA with ResNet50 · CFP-FP | verification accuracy (%): 95.41, 95.47, 95.56, 95.32 | Helped |
| Replace ArcFace with SphereFace using margin 1.35 | Face recognition with CASIA and ResNet50 · AgeDB-30 | verification accuracy (%): 95.15 -> 91.70 | Hurt |
| Replace ArcFace with CosFace using margin 0.35 | Face recognition with CASIA and ResNet50 · AgeDB-30 | verification accuracy (%): 95.15 -> 94.56 | Hurt |
| Add Intra-Loss to Norm-Softmax (ablation) | Norm-Softmax on CASIA with ResNet50 · CFP-FP | verification accuracy (%): 89.79 -> 93.81 | Helped |
| Add Inter-Loss to Norm-Softmax (ablation) | Norm-Softmax on CASIA with ResNet50 · CFP-FP | verification accuracy (%): 89.79 -> 90.67 | Helped |
| Add Intra-Loss to ArcFace (ablation) | ArcFace on CASIA with ResNet50 · CFP-FP | verification accuracy (%): 95.56 -> 95.37 | Hurt |
| Add Inter-Loss to ArcFace (ablation) | ArcFace on CASIA with ResNet50 · CFP-FP | verification accuracy (%): 95.56 -> 95.25 | Hurt |
| Add Triplet-Loss to ArcFace (ablation) | ArcFace on CASIA with ResNet50 · AgeDB-30 | verification accuracy (%): 95.15 -> 94.40 | Hurt |
| Add an iQIYI-VID-trained MLP to ArcFace video embeddings (ablation) | ArcFace on iQIYI-VID · iQIYI-VID test set | MAP (%): 79.80 -> 86.40 | Helped |
| Add model ensemble and context features to the MLP-enhanced system (ablation) | ArcFace on iQIYI-VID · iQIYI-VID test set | MAP (%): 86.40 -> 88.65 | Helped |
| Train ArcFace with MS1MV2+Asian rather than MS1MV2 alone (ablation) | ArcFace, ResNet100 · Trillion-Pairs | identification accuracy (@FPR=1e-3) (%): 80.968 -> 84.840 | Helped |
Closest papers by meaning
- SphereFace: Deep Hypersphere Embedding for Face Recognition (CVPR 2017 · similarity 0.90)
- CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018 · similarity 0.90)
- Mis-Classified Vector Guided Softmax Loss for Face Recognition (AAAI 2020 · similarity 0.89)
- AdaFace: Quality Adaptive Margin for Face Recognition (CVPR 2022 · similarity 0.88)
- UniTSFace: Unified Threshold Integrated Sample-to-Sample Loss for Face Recognition (NeurIPS 2023 · similarity 0.88)
- Attributes Shape the Embedding Space of Face Recognition Models (ICML 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.