Research map / Vision-language and multimodal / Human and embodied modeling
Gender Artifacts in Visual Datasets
Nicole Meister, Dora Zhao, Angelina Wang, Vikram V. Ramaswamy, Ruth Fong, 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 |
|---|---|---|---|
| reduce image resolution to 28x28 and grayscale (ablation) | gender artifact model (ResNet-50) · COCO | AUC (%): 61.9 (color) vs 51.9 (grayscale) | Hurt |
| remove background (show person only) (ablation) | gender artifact model · COCO | AUC (%): 93.4 -> 92.7 | No clear effect |
| occlude person with rectangle, keep background (ablation) | gender artifact model · COCO | AUC (%): 93.4 -> 70.8 | Hurt |
| reduce person to bounding box only (no background) (ablation) | gender artifact model · COCO | AUC (%): 93.4 -> 58.0 | Hurt |
| use only mean RGB color features (ablation) | logistic regression · COCO / OpenImages | AUC (%): 58.0 (COCO), 59.1 (OpenImages) | No clear effect |
| use more random objects (10 -> 30) in object-presence classifier (ablation) | logistic regression · COCO | AUC (%): 58.0 -> 66.8 | Improved |
| train on white instead of black occlusion pixels (ablation) | gender artifact model · COCO / OpenImages | AUC (%): largest change 1.5% COCO, 3.5% OpenImages | No clear effect |
| balance the dataset by gender (ablation) | gender artifact model · COCO | AUC (%): 65.6 -> 63.6 (largest change) | No clear effect |
| pretrain on PASS (no people) instead of ImageNet (ablation) | gender artifact model · COCO | AUC (%): 93.4 -> 89.9 (Full) | No clear effect |
| adversarial gender debiasing (fairness through blindness) | Wang et al. U-Net · COCO val | location of removed pixels: 23.2% person vs 76.8% background | Hurt |
Closest papers by meaning
- Gender Artifacts from Art History to Text-to-Image Generation (NeurIPS 2026 · similarity 0.88)
- SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.87)
- Revealing and Reducing Gender Biases in Vision and Language Assistants (VLAs) (ICLR 2025 · similarity 0.87)
- Understanding Bias in Large-Scale Visual Datasets (NeurIPS 2024 · similarity 0.87)
- Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision (AAAI 2024 · similarity 0.85)
- Finetuning Text-to-Image Diffusion Models for Fairness (ICLR 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.
