CamVLM-8B
📑 Paper | 💻 Code | 🤗 Model & Datasets
✨ Model Description
CamVLM-8B is an LVLM with state-of-the-art surveillance video understanding performance among open-source models, fine-tuned from Qwen3-VL-8B-Instruct. It can actively control real-world surveillance cameras to continuously acquire task-relevant fine-grained visual evidence. It is trained with the carefully crafted SFT and RL training recipe proposed in our paper, utilizing our high-quality training datasets CCTV-Anomaly and CamTrack-53K.
📊 Performance
CamVLM-8B achieves state-of-the-art surveillance understanding performance under three settings:
- Conventional full-frame passive viewpoint
- Center passive viewpoint (simulating a local camera-view region within the panoramic observation space)
- Dynamic viewpoint (simulating a local camera-view region within the panoramic observation space)
Please refer to our paper for detailed results.
🚀 Usage
For result reproduction, please refer to the usage instructions in our GitHub repository. To deploy CamVLM-8B on real-world surveillance cameras, the parameterized viewpoint actions predicted by the model can be converted into executable camera commands according to the hardware specifications of the target camera, such as the horizontal/vertical rotation range, field of view, angular resolution, and optical zoom or focal-length range.
📝 Citation
If you find our work helpful for your research and applications, please cite our paper:
@article{zhang2026thinking,
title={Thinking with Cameras: Active Visual Reasoning via Dynamic Viewpoint Control for Surveillance Video Understanding},
author={Zhang, Xiao and Zeng, Wang and Jin, Sheng and Liu, Wentao and Qian, Chen and Kan, Shichao},
journal={arXiv preprint arXiv:2609.06475},
year={2026}
}