DIS

Why we included this project

This ECCV 2022 project tackles dichotomous image segmentation, a task that pulls a foreground object away from its background no matter how varied the object's structure and detail. It comes with the IS-Net model and the manually labeled DIS5K training set of over 5,400 images. A simple inference.py script and a general-use checkpoint let you run segmentation on your own images without retraining, and there's a hosted demo plus a Hugging Face Space if you want to judge output quality before integrating the code. It's a research codebase rather than a turnkey service, but it's one of the more practical options for teams building background-removal or compositing pipelines.