Research map / Label-efficient and robust learning / Recognition and long-tail
AdaCos: Adaptively Scaling Cosine Logits for Effectively Learning Deep Face Representations
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 ArcFace's fixed scale and angular margin with dynamically adapted cosine-logit scaling | face recognition training with ResNet-50 · LFW | average accuracy: 99.45 -> 99.71 | Helped |
| replace a fixed adaptive scale with a dynamic adaptive scale (ablation) | AdaCos with ResNet-50 · LFW | average accuracy: 99.60 -> 99.71 | Helped |
| replace ArcFace with dynamically adapted cosine-logit scaling | face recognition training with ResNet-50 · LFW at 25k iterations | recognition accuracy: 82.43 -> 88.52 | Helped |
| replace a fixed adaptive scale with a dynamic adaptive scale (ablation) | AdaCos with ResNet-50 · LFW at 25k iterations | recognition accuracy: 85.10 -> 88.52 | Helped |
| replace ArcFace with dynamically adapted cosine-logit scaling | face recognition training with ResNet-50 · LFW at 50k iterations | recognition accuracy: 92.37 -> 95.78 | Helped |
| replace a fixed adaptive scale with a dynamic adaptive scale (ablation) | AdaCos with ResNet-50 · LFW at 50k iterations | recognition accuracy: 94.38 -> 95.78 | Helped |
| replace ArcFace with dynamically adapted cosine-logit scaling | face recognition training with ResNet-50 · LFW at 75k iterations | recognition accuracy: 98.78 -> 99.30 | Helped |
| replace a fixed adaptive scale with a dynamic adaptive scale (ablation) | AdaCos with ResNet-50 · LFW at 75k iterations | recognition accuracy: 99.05 -> 99.30 | Helped |
| replace ArcFace with dynamically adapted cosine-logit scaling | face recognition training with Inception-ResNet · MegaFace, 10^6 distractors | recognition accuracy: 96.88% -> 97.41% | Helped |
| replace a fixed adaptive scale with a dynamic adaptive scale (ablation) | AdaCos with Inception-ResNet · MegaFace, 10^6 distractors | recognition accuracy: 96.85% -> 97.41% | Helped |
| replace ArcFace with dynamically adapted cosine-logit scaling | face recognition training with Inception-ResNet · IJB-C 1:1, FAR 10^-7 | TAR: 72.28% -> 74.07% | Helped |
| replace a fixed adaptive scale with a dynamic adaptive scale (ablation) | AdaCos with Inception-ResNet · IJB-C 1:1, FAR 10^-7 | TAR: 72.66% -> 74.07% | Helped |
Closest papers by meaning
- CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018 · similarity 0.91)
- P2SGrad: Refined Gradients for Optimizing Deep Face Models (CVPR 2019 · similarity 0.87)
- Mis-Classified Vector Guided Softmax Loss for Face Recognition (AAAI 2020 · similarity 0.87)
- ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · similarity 0.85)
- CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition (CVPR 2020 · similarity 0.85)
- Killing Two Birds With One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC (CVPR 2022 · 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.