Research map / Video / Video generation
MotionAtlas: A High-Quality Dataset and Benchmark for Dense Motion Captioning
Weisong Liu, Haochen Wang, Gaokuan Gaokuan, Yuhao Wang, Yikang Zhou, Zhongwei Ren, Guangcan Mai, Anran Wang, Yanwei Li, Xiangtai Li, Zhaoxiang Zhang
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 |
|---|---|---|---|
| fine-tune on MotionAtlas-Data | Qwen3-VL-4B · MotionAtlas-Bench single-frame | accuracy: 19.3 -> 27.7 | Improved |
| fine-tune on MotionAtlas-Data | Qwen3-VL-8B · MotionAtlas-Bench single-frame | accuracy: 24.3 -> 31.6 | Improved |
| fine-tune on MotionAtlas-Data | Molmo2-4B · MotionAtlas-Bench single-frame | accuracy: 13.7 -> 21.6 | Improved |
| fine-tune on region-level MotionAtlas-Data | Qwen3-VL-4B · TOMATO | accuracy: 27.4 -> 35.2 | Improved |
| fine-tune on MotionAtlas-Data | Qwen3-VL-4B · FAVOR-Bench | accuracy: 47.0 -> 55.0 | Improved |
| fine-tune on MotionAtlas-Data | Qwen3-VL-8B · NExT-QA | accuracy: 76.9 -> 77.2 | No clear effect |
| add region detail captions with visual cue instead of general captions only (ablation) | Qwen3-VL-4B SFT · TOMATO | accuracy: 28.4 -> 35.2 | Improved |
| add region detail captions (text reference) instead of general captions only (ablation) | Qwen3-VL-4B SFT · FAVOR-Bench | accuracy: 52.2 -> 55.7 | Improved |
| use visual cue instead of text reference for region (ablation) | Qwen3-VL-4B SFT · FAVOR-Bench | accuracy: 55.7 -> 55.0 | Mixed |
| add self-bootstrap refinement (ablation) | MA data pipeline · MotionAtlas-Bench | accuracy: 36.4 -> 39.9 | Improved |
| add full-video caption (ablation) | MA data pipeline · MotionAtlas-Bench | accuracy: 33.2 -> 39.9 | Improved |
| add spatial crop (ablation) | MA data pipeline · MotionAtlas-Bench | accuracy: 32.7 -> 39.9 | Improved |
Closest papers by meaning
- Dense Video Object Captioning from Disjoint Supervision (ICLR 2025 · similarity 0.90)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.89)
- Towards Understanding Camera Motions in Any Video (NeurIPS 2025 · similarity 0.88)
- Progress-Aware Video Frame Captioning (CVPR 2025 · similarity 0.88)
- FAVOR-Bench: A Comprehensive Benchmark for Fine-Grained Video Motion Understanding (NeurIPS 2025 · similarity 0.88)
- PixFoundation 2.0: Do Video Multi-Modal LLMs Use Motion in Visual Grounding? (NeurIPS 2026 · similarity 0.88)
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.
