Research map / Vision-language and multimodal / Human and embodied modeling
Model-Agnostic Gender Debiased Image Captioning
Yusuke Hirota, Yuta Nakashima, Noa Garcia
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 |
|---|---|---|---|
| Apply LIBRA to captioning models | UpDn, OSCAR, and GRIT · MSCOCO validation | LIC: UpDn 4.7 -> 1.5; OSCAR 2.4 -> 0.3; GRIT 3.1 -> 0.7 | Improved |
| Apply LIBRA to captioning models | UpDn, OSCAR, and GRIT · MSCOCO validation | Error: UpDn 5.6 -> 4.5; OSCAR 3.0 -> 4.6; GRIT 3.5 -> 4.1 | Mixed |
| Apply LIBRA to captioning models | UpDn, OSCAR, and GRIT · MSCOCO validation | BiasAmp: UpDn 1.46 -> -2.23; OSCAR 1.78 -> -1.95; GRIT 3.05 -> 1.57 | Improved |
| Apply LIBRA instead of Gender equalizer | NIC+ · MSCOCO validation | LIC: NIC+ 1.4; Equalizer 6.8; LIBRA 0.4 | Improved |
| Apply LIBRA instead of Gender equalizer | NIC+ · MSCOCO validation | Error: NIC+ 14.6; Equalizer 7.8; LIBRA 5.1 | Improved |
| Use T5 and Merged synthetic captions rather than Swap and T5 captions (ablation) | LIBRA on UpDn · MSCOCO validation | Error: 6.2 -> 4.5 | Improved |
| Use T5 and Merged synthetic captions rather than Swap and T5 captions (ablation) | LIBRA on OSCAR · MSCOCO validation | Error: 6.8 -> 4.6 | Improved |
| Train on all synthetic-caption types rather than T5 and Merged (ablation) | LIBRA on UpDn · MSCOCO validation | LIC: 1.5 -> 1.1 | Improved |
| Train on all synthetic-caption types rather than T5 and Merged (ablation) | LIBRA on OSCAR · MSCOCO validation | LIC: 0.3 -> 0 | Improved |
| Train DCG using BCS-biased captions instead of random perturbations (ablation) | LIBRA on UpDn and OSCAR · MSCOCO validation | LIC and Error: UpDn LIC 2.2 -> 1.5, Error 5.9 -> 4.5; OSCAR LIC 2.0 -> 0.3, Error 5.6 -> 4.6 | Improved |
| Apply LIBRA to captioning models | NIC, SAT, FC, Att2in, UpDn, Transformer, OSCAR, ClipCap, and GRIT · MSCOCO validation | CLIPScore | Mixed |
| Apply LIBRA to captioning models | captioning models evaluated with reference-based metrics · MSCOCO validation | BLEU-4: 79.6 -> 47.5 | Mixed |
Closest papers by meaning
- Women also Snowboard: Overcoming Bias in Captioning Models (ECCV 2018 · similarity 0.93)
- Are Models Biased on Text without Gender-related Language? (ICLR 2024 · similarity 0.88)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.86)
- Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.86)
- Revealing and Reducing Gender Biases in Vision and Language Assistants (VLAs) (ICLR 2025 · similarity 0.86)
- Auto-Search and Refinement: An Automated Framework for Gender Bias Mitigation in Large Language Models (NeurIPS 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.
