Research map / Label-efficient and robust learning / Recognition and long-tail
Towards Fairness in Visual Recognition: Effective Strategies for Bias Mitigation
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 |
|---|---|---|---|
| train domain-specific classifiers and combine their decision-boundary logits | ResNet-18 on CIFAR-10S · CIFAR-10S | Mean accuracy (%): 88.5 ± 0.3 -> 92.0±0.1 | Helped |
| use adversarial training with uniform confusion loss | ResNet-18 on CIFAR-10S · CIFAR-10S | Mean accuracy (%): 88.5±0.3 -> 83.8\pm1.1 | Hurt |
| use adversarial training with gradient reversal and projection | ResNet-18 on CIFAR-10S · CIFAR-10S | Mean accuracy (%): 88.5±0.3 -> 84.1\pm1.0 | Hurt |
| balance class-domain frequencies through oversampling | ResNet-18 on CIFAR-10S · CIFAR-10S | Mean accuracy (%): 88.5 ± 0.3 -> 89.1±0.4 | Helped |
| remove data augmentation from oversampling training (ablation) | Oversampling ResNet-18 · CIFAR-10S | Mean accuracy (%): 89.1±0.4 -> 79.2\pm0.8 | Hurt |
| apply prior shift before aggregating joint class-domain predictions (ablation) | DomainDiscrim ResNet-18 · CIFAR-10S | Mean accuracy (%): 87.3\pm0.3 -> 90.3\pm0.5 | Helped |
| use max over domains instead of summing after prior shift (ablation) | DomainDiscrim ResNet-18 · CIFAR-10S | Mean accuracy (%): 90.3\pm0.5 -> 90.3\pm0.5 | No clear effect |
| apply RBA after prior shift (ablation) | DomainDiscrim ResNet-18 · CIFAR-10S | Accuracy (%): 90.3% -> 90.4% | No clear effect |
| use only the known-domain classifier rather than combining domain-specific boundaries (ablation) | DomainIndependent ResNet-18 · CIFAR-10S | Mean accuracy (%): 92.0\pm0.1 -> 88.9\pm0.4 | Hurt |
| sum per-domain probabilities instead of logits (ablation) | DomainIndependent ResNet-50 · CelebA test, 34 attributes | Weighted mAP (%): 76.3 -> 76.0 | Hurt |
| take the maximum per-domain probability instead of summing logits (ablation) | DomainIndependent ResNet-50 · CelebA test, 34 attributes | Weighted mAP (%): 76.3 -> 75.4 | Hurt |
| use only the classifier for the known domain (ablation) | DomainIndependent ResNet-50 · CelebA test, 34 attributes | Weighted mAP (%): 76.3 -> 73.8 | Hurt |
Closest papers by meaning
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.88)
- Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision (AAAI 2024 · similarity 0.88)
- MABR: Multilayer Adversarial Bias Removal Without Prior Bias Knowledge (AAAI 2025 · similarity 0.88)
- Classes Are Not Equal: An Empirical Study on Image Recognition Fairness (CVPR 2024 · similarity 0.86)
- Discovering and Mitigating Visual Biases through Keyword Explanation (CVPR 2024 · similarity 0.86)
- Towards Fairness under Label Bias in Image Segmentation: Impact, Measurement and Mitigation (NeurIPS 2026 · similarity 0.86)
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.