Research map / Label-efficient and robust learning / Recognition and long-tail
Range Loss for Deep Face Recognition With Long-Tailed Training Data
Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, Yu Qiao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 range loss to softmax, trained on full long-tailed set A-0 | VGG-Face · LFW (A-0) | accuracy: 97.87% to 98.63% | Improved |
| add range loss to softmax, trained on full long-tailed set A-0 | VGG-Face · YTF (A-0) | accuracy: 92.0% to 93.50% | Improved |
| keep the tailed data (A-2 to A-0) when training with range loss (ablation) | Range loss + softmax · LFW | accuracy: accuracy increase by 0.18% in LFW and 0.3% in YTF from A-2 to A-0 | Improved |
| keep the tailed data (A-2 to A-0) when training with contrastive loss (ablation) | Contrastive loss + softmax · LFW | accuracy: 98.47% to 98.35% | Hurt |
| keep the tailed data (A-2 to A-0) when training with triplet loss (ablation) | Triplet loss + softmax · LFW | accuracy: 98.40% to 98.10% | Hurt |
| keep the tailed data (A-2 to A-0) when training with center loss (ablation) | Center loss + softmax · LFW | accuracy: 98.57% to 98.22% | Hurt |
| remove half of tailed data (A-0 to A-2) (ablation) | Softmax VGG-Face · LFW | accuracy: 97.87% to 98.25% | Improved |
| remove all tailed data (A-2 to A-4) (ablation) | Softmax VGG-Face · LFW | accuracy: 98.25% to 95.97% | Hurt |
| remove tail identities, 4000 to 1000 identities (B-0 to B-3) (ablation) | Softmax AlexNet · LFW | accuracy: 78.73% to 83.77% | Improved |
| use k=3 instead of k=2 largest ranges in intra-class loss (ablation) | Range loss · LFW | accuracy: slightly improvement with 0.03% for LFW on average but much hard to train | Mixed |
| add FaceNet hard mining to softmax baseline instead of range loss | VGG on long-tailed data · LFW | accuracy: 96.55% in LFW versus 98.13% in LFW | Hurt |
| add range loss to ResNet trained on long-tailed data | ResNet · BLUFR verification (FAR=0.1%) | accuracy: 90.03% to 92.10% | Improved |
Closest papers by meaning
- Unequal-Training for Deep Face Recognition With Long-Tailed Noisy Data (CVPR 2019 · similarity 0.89)
- NPT-Loss: Demystifying Face Recognition Losses With Nearest Proxies Triplet (TPAMI 2022 · similarity 0.87)
- A Survey on Long-Tailed Visual Recognition (IJCV 2022 · similarity 0.86)
- Improving Face Recognition from Hard Samples via Distribution Distillation Loss (ECCV 2020 · similarity 0.86)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.84)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 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.
