VideoScan

The official implementation of VideoScan+, which is a novel online video inference framework that achieves extremely reduced latency and memory usage.

🌟 Highlights

• Ultra-Low Latency: Optimized streaming inference pipeline
• Memory Efficient: Significantly reduced memory footprint compared to traditional approaches
• Online Processing: Real-time video processing capabilities (around 50 serving FPS)
• Easy Integration: Simple integration with existing video processing pipelines

📅 Timeline

[NEW! 2025.08.05]: VideoScan+ official implementation is released! 🚀🚀🚀

🧐 Quick Start

Installation

cd VideoScan
pip install -r requirements.txt

Basic Usage

1. Offline Video Inference:

python infer.py

2. Online Streaming Video Inference:

python stream_infer.py

📝 Citation

If you find this work useful in your research, please consider citing:

@article{li2025videoscan,
  title={VideoScan: Enabling Efficient Streaming Video Understanding via Frame-level Semantic Carriers},
  author={Li, Ruanjun and Tan, Yuedong and Shi, Yuanming and Shao, Jiawei},
  journal={arXiv preprint arXiv:2503.09387},
  year={2025}
}

📄 License

This project is licensed under the Apache 2.0 License.