Scops self supervised co-part segmentation

The document proposes a self-supervised method called SCOPS to learn co-part segmentation without using any semantic labels. It uses three key losses: 1) Geometric Concentration Loss encourages pixels of the same part to be spatially concentrated, 2) Equivariance Loss enforces consistency under random transformations, and 3) Semantic Consistency Loss avoids bias towards background parts by using image-level labels from a pretrained model. Together these losses allow SCOPS to segment objects into semantically consistent parts by leveraging only unlabeled images of the same category.