#174 · Primary category: Computer Vision

BackgroundMattingV2

computer-vision machine-learning matting real-time

Real-Time High-Resolution Background Matting

Project last updated:06/19/24

GitHub Stars

7.2K

Forks

960

Contributors

5

License

MIT

Why we included this project

BackgroundMattingV2 tackles a specific version of the person-separation problem: it expects you to capture one clean frame of the background ahead of time, then uses that reference to pull a soft alpha matte at 4K 30fps or HD 60fps on a modern GPU. That precondition matters, because it rules out scenes you cannot control, but for controlled setups like webcam streams or studio shots it is a strong, well-documented baseline. The authors released the VideoMatte240K and PhotoMatte85 datasets alongside the code, so teams can retrain or fine-tune on the same data behind the CVPR 2021 paper. Be aware the shipped scripts measure tensor throughput rather than end-to-end video speed; you will need to add hardware encoding and parallel frame loading for real use. If you want to judge matting quality before committing that engineering, the weights and usage docs are a reasonable place to start.

Articles for this project

No articles for this project yet.

To suggest a topic or contribute an article, contact us.

Related projects in this category