Research map / Label-efficient and robust learning / Recognition and long-tail
SphereFace: Deep Hypersphere Embedding for 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 standard softmax supervision with A-Softmax angular margin (m=4) | 64-layer CNN · LFW | accuracy: 98.20% to 99.42% | Helped |
| replace standard softmax supervision with A-Softmax angular margin (m=4) | 64-layer CNN · YTF | accuracy: 93.4% to 95.0% | Helped |
| increase angular margin parameter m (ablation) | A-Softmax · LFW and YTF | accuracy: LFW: 97.90 -> 99.42; YTF: 93.2 -> 95.0 | Helped |
| remove the last ReLU (ablation) | 4-layer CNN with SphereFace modifications · LFW | accuracy: 95.13 -> 96.37 | Helped |
| zero the final fully connected layer biases (ablation) | 4-layer CNN with last ReLU removed · LFW | accuracy: 96.37 -> 96.40 | Helped |
| normalize the final fully connected layer weights (ablation) | 4-layer CNN with last ReLU and biases removed · LFW | accuracy: 96.40 -> 96.63 | Helped |
| add A-Softmax angular margin with m=2 (ablation) | 4-layer CNN with normalized weights · LFW | accuracy: 96.63 -> 97.67 | Helped |
| increase A-Softmax margin from m=2 to m=4 (ablation) | 4-layer CNN · LFW | Angular Fisher Score: 0.2277 -> 0.1709 | Helped |
| increase CNN depth (ablation) | SphereFace with A-Softmax m=4 · MegaFace Challenge 1 small protocol | Rank-1 identification accuracy: 57.529 -> 72.729 | Helped |
| use 3-patch feature concatenation ensemble instead of a single model (ablation) | SphereFace · MegaFace Challenge 1 small protocol | Rank-1 identification accuracy: 72.729 -> 75.766 | Helped |
| remove the last ReLU (ablation) | SphereFace feature network · feature learning | feasible angular learning space | Helped |
| normalize class weights (ablation) | SphereFace · face verification under training-class imbalance | verification performance | Helped |
Closest papers by meaning
- ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · similarity 0.90)
- SphereFace2: Binary Classification is All You Need for Deep Face Recognition (ICLR 2022 · similarity 0.89)
- UniformFace: Learning Deep Equidistributed Representation for Face Recognition (CVPR 2019 · similarity 0.88)
- Softmax Dissection: Towards Understanding Intra- and Inter-Class Objective for Embedding Learning (AAAI 2020 · similarity 0.88)
- CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018 · similarity 0.87)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (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.