STEGO

Why we included this project

Pixel-level labels are expensive to produce, and for many image domains they simply don't exist. STEGO attacks that problem directly: it learns semantic segmentation without any human-annotated masks, using the official ICLR 2022 implementation that distills feature correspondences from a self-supervised vision transformer. Pretrained models, evaluation scripts, and a Colab demo are all in the repo, so you can see it segment Cityscapes or COCO-Stuff scenes before committing to training. The README also walks through bringing your own unlabeled dataset, which matters when a dense label set for your domain would be impractical to build. If you're working on unsupervised or weakly supervised computer vision, this is a well-documented reference worth studying and adapting.