申請存取 · Request access
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
研究用模型權重，僅限非商業研究（CC-BY-NC-4.0），需人工審批。
技術細節、使用說明與權重檔案於申請獲批後可見。
Research weights for non-commercial research only (CC-BY-NC-4.0), approved
manually. Technical details, usage instructions and the weight files are visible once approved.
Log in or Sign Up to review the conditions and access this model content.
GeoMo-Lite: a compact camera-motion recogniser
This repository holds two camera-motion recognisers, GeoMo-Lite and GeoMo. They recognise camera motion in video clips, for example pan, tilt, truck, pedestal, push/pull, zoom, roll, arc, tracking and static camera, including when and in which order movements happen within a shot.
Results
All numbers are accuracy (%), mean of 3 training seeds. Each test set was scored once.
| Benchmark | GeoMo-Lite | GeoMo | Best published model we compare with |
|---|---|---|---|
| ShotBench (camera movement, 464 Qs) | 87.8 | 87.0 | 59.7 (ShotVL-7B) |
| ACaM, real videos (1,385 Qs) | 73.1 | 74.5 | 72.8 (ACaM-SFT-8B) |
| ACaM, synthetic videos (1,179 Qs) | 89.5 | 89.6 | 78.2 (ACaM-SFT-8B) |
| CameraBench-Binary, 15 primitives (7,679 Qs) | 92.6 | 90.0 | 80.6 (Cam-Motion-72B) |
| CJ-Synth matched-twin test (936 Qs) | 87.9 | 92.5 | 65.3 (CamDistill-8B) |
Human accuracy on ACaM (from the ACaM paper) is 93.4 (real) / 98.1 (synthetic); both models are still below human level there.
Access
Weights, usage instructions and technical details are available after your access request is approved.
License
CC BY-NC 4.0, non-commercial research use only. Some of the preprocessing tools used with this model carry their own non-commercial licences.
Citation
@misc{lin2026geomolite,
title = {GeoMo-Lite: a compact camera-motion recogniser},
author = {Zeteng Lin and Jing Tang},
year = {2026},
note = {Hugging Face model repository}
}