Research map / Vision-language and multimodal / Human and embodied modeling
Women also Snowboard: Overcoming Bias in Captioning Models
Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, Anna Rohrbach
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 Appearance Confusion Loss to Equalizer w/o ACL (ablation) | Equalizer · MSCOCO-Bias | gender error rate: 7.57 -> 7.02 | Improved |
| add Confident Loss to Equalizer w/o Conf (ablation) | Equalizer · MSCOCO-Bias | gender error rate: 9.62 -> 7.02 | Improved |
| add Appearance Confusion Loss to Equalizer w/o ACL (ablation) | Equalizer · MSCOCO-Balanced | gender error rate: 10.10 -> 8.10 | Improved |
| add Confident Loss to Equalizer w/o Conf (ablation) | Equalizer · MSCOCO-Balanced | gender error rate: 13.90 -> 8.10 | Improved |
| add Confident Loss to Equalizer w/o Conf (ablation) | Equalizer · MSCOCO-Bias | gender ratio difference: 0.09 -> -0.03 | Improved |
| add Appearance Confusion Loss to Equalizer w/o ACL (ablation) | Equalizer · MSCOCO-Bias | outcome divergence between genders: 0.031 -> 0.018 | Improved |
| add Confident Loss to Equalizer w/o Conf (ablation) | Equalizer · MSCOCO-Bias | outcome divergence between genders: 0.075 -> 0.018 | Improved |
| add Appearance Confusion Loss to Equalizer w/o ACL (ablation) | Equalizer · MSCOCO-Balanced pointing game (Grad-CAM) | pointing accuracy, all: 43.8 -> 47.5 | Improved |
| add Appearance Confusion Loss to Equalizer w/o ACL (ablation) | Equalizer · masked images, man-to-woman ratio | predicted gender ratio (ideal 1.0): 3.45 -> 1.98 | Improved |
| rebalance training data by resampling women images (Balanced) vs fine-tuning | Baseline-FT · MSCOCO-Bias | gender error rate: 12.83 -> 12.85 | No clear effect |
| upweight gender words in cross entropy loss by 10 | Baseline-FT · MSCOCO-Bias vs MSCOCO-Balanced | gender error rate: 12.83 -> 13.56 (Bias); 19.30 -> 16.30 (Balanced) | Mixed |
| replace baseline with Equalizer (both losses) on shifted test distribution | Baseline-FT · MSCOCO-Balanced | gender error rate: 19.30 -> 8.10 | Improved |
Closest papers by meaning
- Model-Agnostic Gender Debiased Image Captioning (CVPR 2023 · similarity 0.93)
- VisoGender: A dataset for benchmarking gender bias in image-text pronoun resolution (NeurIPS 2023 · similarity 0.89)
- Understanding and Evaluating Racial Biases in Image Captioning (ICCV 2021 · similarity 0.89)
- Are Models Biased on Text without Gender-related Language? (ICLR 2024 · similarity 0.86)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.86)
- Revealing and Reducing Gender Biases in Vision and Language Assistants (VLAs) (ICLR 2025 · 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.
