Research map / Video / Video-language understanding
MotionChain: Fine-Grained Video Motion Understanding via Structured Decomposition
Hao Yang, Bo Xu, Jun Dan, Sijia Chen, Baigui Sun, Yang Liu
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 MotionChain to Qwen3-VL-8B | Qwen3-VL-8B · FAVOR-Bench | overall accuracy (%): 56.68 -> 58.03 | Improved |
| apply MotionChain to Qwen3-VL-8B | Qwen3-VL-8B · MotionBench dev | overall accuracy (%): 62.87 -> 63.59 | Improved |
| apply MotionChain to InternVL3.5-8B | InternVL3.5-8B · MotionBench dev | overall accuracy (%): 56.25 -> 57.39 | Improved |
| add optical-flow textual cues for camera motion | InternVL3.5-8B · MotionBench dev | Camera Motion accuracy (%): 48.57 -> 54.29 | Improved |
| add optical-flow camera cues (dedicated tool) (ablation) | Qwen3-VL-8B · FAVOR-Bench | Camera Motion accuracy (%): 46.05 -> 49.02 | Improved |
| add visual enhancement (temporal annotation + SAM3 ROI) on top of MC Frame sampling (ablation) | MotionChain (Qwen3-VL-8B) · FAVOR-Bench | overall accuracy (%): 58.03 -> 57.93 | No clear effect |
| add MotionChain boundary-aware frame sampling (ablation) | Qwen3-VL-8B · FAVOR-Bench | overall accuracy (%): 56.68 -> 58.03 | Improved |
| use MotionChain instead of 3-VLM majority vote (ablation) | MotionBench pipeline · MotionBench dev | overall accuracy (%): 58.51 -> 63.59 | Improved |
| use MotionChain instead of larger uniform frame budgets (ablation) | Qwen3-VL-8B · MotionBench dev | overall accuracy (%): 61.30 -> 63.59 | Improved |
| increase uniform frames to 128 at lower resolution (ablation) | Qwen3-VL-8B · MotionBench dev | overall accuracy (%): 60.85 -> 60.35 | No clear effect |
| run full MotionChain pipeline instead of direct answering | Qwen3-VL-8B · MotionBench dev | median latency (s): 3.95s -> 26.97s | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Open-o3-Video (7B) — Open-o3-Video: Grounded Video Reasoning with Explicit Spatio-Temporal Evidence (ICML 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Motion-Grounded Video Reasoning: Understanding and Perceiving Motion at Pixel Level (CVPR 2025 · similarity 0.90)
- FAVOR-Bench: A Comprehensive Benchmark for Fine-Grained Video Motion Understanding (NeurIPS 2025 · similarity 0.90)
- Compositional 4D Dynamic Scenes Understanding with Physics Priors for Video Question Answering (ICLR 2025 · similarity 0.89)
- Towards Understanding Camera Motions in Any Video (NeurIPS 2025 · similarity 0.88)
- Video-of-Thought: Step-by-Step Video Reasoning from Perception to Cognition (ICML 2024 · similarity 0.88)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · 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.
