Eran Treister

We study label-free semantic segmentation in a strict from-scratch setting, where neither pixel annotations nor external pretrained backbones are available. We propose SPARC, an end-to-end framework that learns a differentiable superpixel tokenization, performs graph message passing over superpixel embeddings, and decodes the refined features back to a pixel-wise segmentation map. The superpixel assignment is predicted as soft, differentiable correspondences, enabling joint optimization of (i) the tokenization, (ii) long-range graph reasoning, and (iii) a segmentation decoder. Training uses mutual-information-based consistency objectives together with superpixel regularizers that encourage spatial smoothness, boundary adherence, and faithful reconstruction. On COCO-Stuff and ISPRS Potsdam, SPARC improves over recent label-free baselines under the same training protocol, while operating on a compact graph that reduces the cost of long-range interactions.