Research map / Vision-language and multimodal / Human and embodied modeling
REVISE: A Tool for Measuring and Mitigating Bias in Visual Datasets
Angelina Wang, Arvind Narayanan, Olga Russakovsky
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 |
|---|---|---|---|
| male instead of female person in COCO image area | COCO person annotations · COCO person prominence | fraction of image area: .138 -> .268 | Improved |
| male instead of female person, distance from center | COCO person annotations · COCO person prominence | distance from image center: .510 -> .363 | Improved |
| darker skin tone of person | COCO person annotations · COCO skin tone trend | person size: p = 2.11e-7 | Hurt |
| darker skin tone of person | COCO person annotations · COCO skin tone trend | distance from center: p = .014 | Hurt |
| same-class vs different-class same-supercategory object pairs in feature space | ResNet-110 CIFAR-10 features · COCO appearance feature distance | mean feature distance: 5.91 -> 6.24 | Improved |
| unrelated object pairs vs same-supercategory pairs in feature space | ResNet-110 CIFAR-10 features · COCO appearance feature distance | mean feature distance: 6.24 -> 6.48 | Improved |
| add word-embedding supercategory binning | OpenImages labels · OpenImages to COCO supercategories | appropriate mappings of 50: 44 of 50 | Improved |
| darker skin tone of person, regional origin of images | COCO images of people · COCO geography by people | share from North America and Europe (%): 85.5 -> 58.2 | Hurt |
| darker skin tone of person, urban location | COCO US images · COCO US urban share | share in urban area (%): 84.4 -> 92.7 | Improved |
| New York City instead of Northern California | BDD100K · BDD100K weather | snowy image share (%): 0.3 -> 10 | Improved |
| unidentifiable gender person in COCO labeled male by annotators | COCO gender labels · COCO unidentifiable gender | share labeled male (%): 77 | Hurt |
| unidentifiable gender person in OpenImages labeled male by annotators | OpenImages gender labels · OpenImages unidentifiable gender | share labeled male (%): 69 | Hurt |
Closest papers by meaning
- Understanding Bias in Large-Scale Visual Datasets (NeurIPS 2024 · similarity 0.88)
- Evaluating Model Bias Requires Characterizing its Mistakes (ICML 2024 · similarity 0.84)
- OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes (ICLR 2025 · similarity 0.84)
- Discovering and Mitigating Visual Biases through Keyword Explanation (CVPR 2024 · similarity 0.84)
- VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations (ICLR 2026 · similarity 0.83)
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.83)
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.
