Qwen2-VL-7B-Instruct-Traffic

Qwen2-VL-7B-Instruct-Traffic is a multimodal model fine-tuned on the MITS (Multimodal Intelligent Traffic Surveillance) dataset for intelligent traffic surveillance scenarios.

• Tasks: recognition, counting, localization, background awareness, reasoning
• Data: 170,400 images + ~5M instruction-following VQA pairs from MITS
• Modality: Image + Text → Text
• Domain: traffic scenes (congestion, accidents, construction, smoke/fireworks, unusual weather, spills, etc.)

Quick Links

• 📚 Dataset: zhaokaikai/Multimodal_Intelligent_Traffic_Surveillance
• 💻 Usage & examples: please refer to the GitHub repo
  https://github.com/LifeIsSoSolong/Multimodal-Intelligent-Traffic-Surveillance-Dataset-Models

Intended Use

• Urban traffic monitoring, incident analysis, visual question answering for transportation management
• Research on ITS-specific multimodal reasoning and instruction following

Model Inputs/Outputs

• Input: an image (traffic scene) + a natural language instruction/question
• Output: a natural language response (e.g., description, count, event reasoning)

Training Summary

• Objective: instruction tuning on MITS traffic QA
• Backbone family: Qwen2-VL 7B Instruct
• Notes: align vision-language features to traffic-centric concepts and events

Limitations & Notes

• The model may make mistakes on rare objects or extreme weather/night scenes not well represented in training.
• Not a safety-critical system; human verification is required for real-world decisions.

License

• Follow the licenses of this model and the MITS dataset as stated on their ModelScope pages.

Citation

If you use this model or dataset, please cite:

@article{zhao2025mits,
  title   = {MITS: A large-scale multimodal benchmark dataset for Intelligent Traffic Surveillance},
  author  = {Zhao, Kaikai and Liu, Zhaoxiang and Wang, Peng and Wang, Xin and Ma, Zhicheng and Xu, Yajun and Zhang, Wenjing and Nan, Yibing and Wang, Kai and Lian, Shiguo},
  journal = {Image and Vision Computing},
  pages   = {105736},
  year    = {2025},
  publisher = {Elsevier}
}

Contact

Unicom AI