Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Imbalanced Learning for Face Recognition and Attribute Prediction
Chen Huang, Yining Li, Chen Change Loy, Xiaoou Tang
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 uniform-to-hardest query cluster sampling (choose highest-loss cluster instead of uniform) (ablation) | CLMLE · LFW | accuracy (%): 99.52 -> 99.62 | Improved |
| add highest-loss query cluster sampling instead of uniform (ablation) | CLMLE · CelebA | average balanced accuracy (%): 88.23 -> 88.78 | Improved |
| add cost-sensitive learning (inverse class frequency in batch) (ablation) | CLMLE · LFW | accuracy (%): 99.57 -> 99.62 | Improved |
| add cost-sensitive learning (inverse class frequency in batch) (ablation) | CLMLE · CelebA | average balanced accuracy (%): 87.46 -> 88.78 | Improved |
| use k-nearest cluster classifier instead of instance-wise kNN (ablation) | CLMLE · CelebA | average balanced accuracy (%): 88.59 -> 88.78 | Improved |
| replace prior features for initial clustering (softmax vs triplet pretraining) (ablation) | CLMLE · LFW | accuracy (%): no more than 0.02% | No clear effect |
| replace prior features for initial clustering (face recognition vs multi-attribute softmax) (ablation) | CLMLE · CelebA | accuracy (%): no more than 0.3% | No clear effect |
| replace quintuplet sampling (LMLE) with entire cluster sampling (CLMLE) | LMLE · CelebA | average balanced accuracy (%): 83.83 -> 88.78 | Improved |
| replace quintuplet sampling (LMLE) with cluster sampling (CLMLE) | LMLE · MegaFace Challenge 1 (small protocol) | verification TAR (%): 89.45 -> 91.85 | Improved |
| use CLMLE instead of CE+CRL on imbalanced CIFAR-100 | CE+CRL · CIFAR-100 (gamma=0.5) | mean class-balanced accuracy (%): 37.4 -> 39.7 | Improved |
| use CLMLE instead of CE+CRL on imbalanced MNIST-rot-back-image | CE+CRL · MNIST-rot-back-image (gamma=0.5) | mean class-balanced accuracy (%): 64.5 -> 68.8 | Improved |
| use CLMLE instead of Triplet+ on imbalanced CIFAR-100 | Triplet+ · CIFAR-100 (gamma=0.5) | mean class-balanced accuracy (%): 32.2 -> 39.7 | Improved |
Closest papers by meaning
- Learning Deep Representation for Imbalanced Classification (CVPR 2016 · similarity 0.92)
- Class Rectification Hard Mining for Imbalanced Deep Learning (ICCV 2017 · similarity 0.89)
- Balanced Data, Imbalanced Spectra: Unveiling Class Disparities with Spectral Imbalance (ICML 2024 · similarity 0.86)
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.85)
- Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.85)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · 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.
