Research map / Video / Video-language understanding
MotionSight: Boosting Fine-Grained Motion Understanding in Multimodal LLMs
Yipeng Du, Tiehan Fan, Kepan Nan, Rui Xie, Penghao Zhou, Xiang Li, Jian Yang, Zhenheng Yang, Ying Tai
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 visual spotlight (darken background outside tracked boxes) (ablation) | Qwen2.5VL-7B · MotionBench object motion | OM AVG accuracy: 51.7 -> 53.0 | Improved |
| use background blur visual prompt (ablation) | Qwen2.5VL-7B · MotionBench object motion | OM AVG accuracy: 51.7 -> 49.3 | Hurt |
| add object crop prompt (ablation) | Qwen2.5VL-7B · MotionBench object motion | OM AVG accuracy: 51.7 -> 52.5 | No clear effect |
| add pose estimation prompt (ablation) | Qwen2.5VL-7B · MotionBench object motion | OM AVG accuracy: 51.7 -> 50.6 | Hurt |
| add global motion blur for camera motion (ablation) | Qwen2.5VL-7B · MotionBench camera motion | accuracy: 34.0 -> 48.3 | Improved |
| increase motion blur temporal window from 7 to 11 (ablation) | MotionSight motion blur · MotionBench camera motion | accuracy: 48.3 -> 41.0 | Hurt |
| apply MotionSight to InternVL3-78B (ablation) | InternVL3-78B · MotionBench | overall accuracy: 61.5 -> 63.0 | Improved |
| apply MotionSight to Qwen2.5VL-7B | Qwen2.5VL-7B · VideoMME | overall accuracy: 73.7 -> 76.0 | Improved |
| apply MotionSight to Qwen2.5VL-7B | Qwen2.5VL-7B · TempCompass captioning | average accuracy: 54.6 -> 64.0 | Improved |
| fine-tune Wan2.1-T2V-1.3B on MotionVid-QA (ablation) | Wan2.1-T2V-1.3B · VBench | Human Action: 64.2 -> 71.0 | Improved |
| sweep background darkening factor (ablation) | MotionSight spotlight · MotionBench OM AVG | accuracy: 51.9 to 53.0 across 0.1-1.0 | No clear effect |
Closest papers by meaning
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.90)
- FAVOR-Bench: A Comprehensive Benchmark for Fine-Grained Video Motion Understanding (NeurIPS 2025 · similarity 0.90)
- PixFoundation 2.0: Do Video Multi-Modal LLMs Use Motion in Visual Grounding? (NeurIPS 2026 · similarity 0.89)
- Improving LLM Video Understanding with 16 Frames Per Second (ICML 2025 · similarity 0.89)
- Which Way Did It Move? Diagnosing and Overcoming Directional Motion Blindness in Video LLMs (NeurIPS 2026 · similarity 0.88)
- MG-MotionLLM: A Unified Framework for Motion Comprehension and Generation across Multiple Granularities (CVPR 2025 · 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.
