ECCV 2026

MotionAtlas

Detailed Region Captioning for Motion-Centric Videos

Weisong Liu1*, Haochen Wang1*, Kuan Gao2*, Yuhao Wang4, Yikang Zhou5, Zhongwei Ren6, Jacky Mai3, Anna Wang3, Yanwei Li2, Jason Li3‡, Zhaoxiang Zhang1†

1CASIA    2SJTU    3NTU    4PKU    5WHU    6BJTU

* Equal contribution. Project lead. Corresponding author.

TL;DR: MotionAtlas shifts motion captioning from global video descriptions to region-aware motion captions, enabling precise evaluation through MotionAtlas-Bench and scalable training through MotionAtlas-Data.

Intro

MotionAtlas shifts motion captioning from global video descriptions to region-aware motion captions, with scalable training data and checklist-style evaluation over referred objects.

Benchmark Results

Consistent Gains Across Benchmarks

Our MotionAtlas consistently outperforms Qwen3-VL baselines across eight motion-related benchmarks.

Quantitative comparison between MotionAtlas and Qwen3-VL baselines across eight benchmarks
Evaluation

MotionAtlas-Bench

Each video is decomposed into events and checked through multiple-choice questions over temporal cues, kinematics, references, and local regions.

MotionAtlas teaser showing region-aware motion captioning and checklist-style evaluation

Demos

Examples from MotionAtlas-Data. Given a user-specified region (highlighted in each clip), MotionAtlas produces a detailed, temporally grounded description of that region. Use the arrows to browse examples.

2,073

Fine-Grained MCQs

MotionAtlas-Bench uses dense checklist-style questions to judge detailed motion captions over referred objects.

159K

Training Samples

MotionAtlas-Data provides scalable region-level motion captions refined to suppress fine-grained hallucinations.

23

Verbs / Sample

MotionAtlas-Data emphasizes dense action verbs and detailed temporal motion descriptions.

Results

Training on MotionAtlas-Data improves both region-level motion captioning and broader motion-related video understanding.

Table 3

Main Results on MotionAtlas-Bench

Model SF Overall SF Parts SF Kin. FS Overall FS Parts FS Kin.
Gemini 3 Pro36.434.732.036.533.538.1
GPT-5.236.934.034.237.638.836.6
Qwen3-VL-235B30.527.828.933.733.231.1
Qwen3-VL-4B19.320.014.121.722.416.5
+ MotionAtlas-Data27.7 ↑ 8.427.926.930.1 ↑ 8.430.329.3
Qwen3-VL-8B24.323.920.326.724.626.7
+ MotionAtlas-Data31.6 ↑ 7.331.230.634.1 ↑ 7.433.633.0

SF = Single-Frame Grounding, FS = Full-Sequence Grounding. Values are accuracy.

Table 4

Motion-Related Video Understanding

Model MotionBench DREAM-1K TOMATO NExT-QA TempCompass FAVOR TVBench
GPT-5.265.442.253.079.973.056.853.8
Gemini 2.5 Pro62.042.748.679.873.758.859.9
Qwen3-VL-4B55.935.627.471.669.647.047.2
+ MotionAtlas-Data61.9 ↑ 6.038.9 ↑ 3.335.2 ↑ 7.874.0 ↑ 2.474.2 ↑ 4.655.0 ↑ 8.151.2 ↑ 4.0
Qwen3-VL-8B59.038.734.076.971.854.151.4
+ MotionAtlas-Data62.6 ↑ 3.639.6 ↑ 0.936.5 ↑ 2.577.2 ↑ 0.275.1 ↑ 3.357.7 ↑ 3.652.9 ↑ 1.5
Table 5

Training Data Scale

Method MotionAtlas MotionBench DREAM-1K TOMATO NExT-QA TempCompass FAVOR TVBench
Qwen3-VL-4B19.255.935.927.471.669.647.047.2
w/ MotionAtlas-Data
20% (32K)22.958.936.928.472.271.248.147.0
60% (95K)24.659.537.030.173.072.350.949.0
100% (159K)28.361.938.935.274.074.255.051.2
w/o MotionAtlas-Data
20% (32K)12.957.437.329.070.970.847.446.7
60% (95K)12.958.836.930.771.372.350.647.4
100% (159K)12.260.538.328.471.973.352.248.5
Fig. 3

Data Scaling Curve

Data scaling curve Average benchmark score rises more sharply when training includes MotionAtlas-Data. Baseline 50.7 0% 20% 60% 100% 55.8 53.3 w/ MotionAtlas-Data w/o MotionAtlas-Data

Adding MotionAtlas-Data brings more significant improvements as training data scales.

Table 6

Pipeline Ablation

MethodAccRecallPrecision
MA Pipeline (full)39.968.258.5
w/o Self-Bootstrap36.464.156.8
w/o Full-Video Caption33.258.956.4
w/o Spatial Crop32.760.953.6

Each component contributes to more accurate and recall-rich motion captions.

BibTeX

@article{liu2026motionatlas,
  title={MotionAtlas: Detailed Region Captioning for Motion-Centric Videos},
  author={Liu, Weisong and Wang, Haochen and Gao, Kuan and Wang, Yuhao and Zhou, Yikang and Ren, Zhongwei and Mai, Jacky and Wang, Anna and Li, Yanwei and Li, Jason and Zhang, Zhaoxiang},
  journal={arXiv preprint arXiv:2606.29531},
  year={2026},
  eprint={2606.29531},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2606.29531}
}