Research map / Video / Video-language understanding
Reverse Distribution Based Video Moment Retrieval for Effective Bias Elimination
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Defines data and model bias in video moment retrieval, proposes ReDis-VMR, a reverse-distribution test-set generation protocol with a bias-ratio metric, and the DEA training pipeline (expandable structure plus fair loss).
Problem
Video moment retrieval models can score well while ignoring the video input because of concentrated temporal distributions in training/test data (data bias) and model candidate tendencies (model bias); existing resplitting evaluation addresses only data bias and does not quantify bias.
Approach
Generate model-specific reverse-distribution train/test data from the model's own prediction density via Gaussian kernel estimation, quantify residual bias with a KL-based bias ratio, and train with dynamic expandable model structure plus a fair loss pushing predicted confidence toward uniform.
Key results
- On Charades-STA under ReDis-VMR, ViGA+DEA reaches R1@0.5 of 26.28 and mIoU 29.61 versus ViGA's 16.75/... (ViGA R1@0.5 16.75, mIoU 21.94 on Charades-STA ReDis setting).
- ReDis-VMR lowers Charades-STA ViGA bias ratio from 93.04 (traditional) to 36.83, while resplitting gives 82.15.
- Table 3 reports CLIP-based bias ratio falling from 88.55 to 57.22 under ReDis, and DEA further reduces it to 47.41.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DEA to ViGA | ViGA · Charades-STA, ReDis-VMR | mIoU: 21.94 -> 29.61 | Helped |
| Add DEA to Vslnet | Vslnet · Charades-STA, ReDis-VMR | mIoU: 26.76 -> 35.13 | Helped |
| Add DEA to Shuffle | Shuffle · Charades-STA, ReDis-VMR | mIoU: 27.81 -> 41.66 | Helped |
| Add DEA to ViGA | ViGA · ActivityNet Captions, ReDis-VMR | mIoU: 27.87 -> 29.66 | Helped |
| Add DEA to Vslnet | Vslnet · ActivityNet Captions, ReDis-VMR | mIoU: 25.58 -> 31.02 | Helped |
| Add DEA to Shuffle | Shuffle · ActivityNet Captions, ReDis-VMR | mIoU: 31.84 -> 37.09 | Helped |
| Add dynamic model structure only (ablation) | ViGA · Charades-STA, ReDis-VMR | R1@0.3: 33.76 -> 34.23 | Helped |
| Add dynamic model structure only (ablation) | ViGA · Charades-STA, ReDis-VMR | R1@0.5: 17.38 -> 16.71 | Hurt |
| Add replay mechanism only (ablation) | ViGA · Charades-STA, ReDis-VMR | mIoU: 21.94 -> 28.87 | Helped |
| Add fair loss only (ablation) | ViGA · Charades-STA, ReDis-VMR | mIoU: 21.94 -> 21.94 | No clear effect |
| Add dynamic model structure and replay mechanism (ablation) | ViGA · Charades-STA, ReDis-VMR | mIoU: 21.94 -> 29.40 | Helped |
| Add dynamic model structure and fair loss (ablation) | ViGA · Charades-STA, ReDis-VMR | mIoU: 21.94 -> 28.89 | Helped |
Surprising findings
- Resplitting VMR is ineffective at removing data bias and in some cases increases the model bias ratio compared with traditional evaluation.
Limitations
- Bias ratio depends on the KL-based estimate and kernel bandwidth choices; the framework is evaluated only on a few baselines and two main datasets.
- The fair loss weight lambda must be tuned, and the best value (1) is reported only from a sensitivity study on one baseline.
Experimental setup
Datasets: Charades-STA ActivityNet Captions HiREST QVHighlights
Benchmarks: Charades-STA ActivityNet Captions HiREST QVHighlights
Metrics: R1@0.3 R1@0.5 R1@0.7 mIoU bias ratio
Baselines: ViGA VSLNet Shuffle Resplitting VMR Traditional VMR
Builds on: resplitting (Yuan et al. 2021) ViGA VSLNet Shuffle (Hao et al. 2022)
Closest papers by meaning
- Fewer Steps, Better Performance: Efficient Cross-Modal Clip Trimming for Video Moment Retrieval Using Language (AAAI 2024 · similarity 0.87)
- When One Moment Isn't Enough: Multi-Moment Retrieval with Cross-Moment Interactions (NeurIPS 2025 · similarity 0.87)
- Generative Video Diffusion for Unseen Novel Semantic Video Moment Retrieval (AAAI 2025 · similarity 0.85)
- VERIFIED: A Video Corpus Moment Retrieval Benchmark for Fine-Grained Video Understanding (NeurIPS 2024 · similarity 0.84)
- MomentSeeker: A Task-Oriented Benchmark For Long-Video Moment Retrieval (NeurIPS 2025 · similarity 0.84)
- Unbiasing through Textual Descriptions: Mitigating Representation Bias in Video Benchmarks (CVPR 2025 · similarity 0.84)
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.