Research map / Video / Video-language understanding
Video DataFlywheel: Resolving the Impossible Data Trinity in Video-Language Understanding
Xiao Wang, Jianlong Wu, Zijia Lin, Fuzheng Zhang, Di Zhang, Liqiang Nie
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 |
|---|---|---|---|
| replace baseline refined datasets with VidDF-refined dataset | video-LLM pre-training · PT caption eval (770K) | mean CIDEr: 90.1 -> 92.9 | Improved |
| use AdaTaiLr noise control instead of no noise control (ablation) | VidDF PT · InternVid PT eval | mean CIDEr: 90.1 -> 92.0 | Improved |
| use AdaTaiLr instead of TaiLr (ablation) | noise control · InternVid PT eval | mean CIDEr: 90.9 -> 92.0 | Improved |
| add CTPR noise control (ablation) | InternVid PT · InternVid PT eval | mean CIDEr: 90.1 -> 88.6 | Hurt |
| add Loss Truncation (ablation) | InternVid PT · InternVid PT eval | mean CIDEr: 90.1 -> 89.8 | Hurt |
| add similarity filtering (ablation) | InternVid PT · InternVid PT eval | mean CIDEr: 90.1 -> 90.5 | Mixed |
| iterative refinement alone when scaling data (ablation) | Base + iterative refine · PT eval | CIDEr | Hurt |
| add AdaTaiLr to iterative refinement (ablation) | VidDF · PT eval scaling | CIDEr | Improved |
| use very small PT data (85K) with VidDF variants (ablation) | VidDF · PT eval 85K | CIDEr | Hurt |
| increase lambda in AdaTaiLr (ablation) | AdaTaiLr · PT eval | CIDEr | Mixed |
| scale PT data from 85K to 2310K | PT stage · video caption / QA | mean CIDEr / QA acc: 84.7 -> 92.6 | Improved |
| use VidDF dataset instead of Panda70M subset for retrieval | UMT-L · MSRVTT R@1 | R@1: 58.4 -> 56.1 | Hurt |
Closest papers by meaning
- Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (ICLR 2025 · similarity 0.90)
- Unleashing Hour-Scale Video Training for Long Video-Language Understanding (NeurIPS 2025 · similarity 0.88)
- HourVideo: 1-Hour Video-Language Understanding (NeurIPS 2024 · similarity 0.88)
- SF2T: Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding (CVPR 2025 · similarity 0.88)
- E.T. Bench: Towards Open-Ended Event-Level Video-Language Understanding (NeurIPS 2024 · similarity 0.87)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.87)
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.
