Rethinking Self-Supervised Semantic Segmentation: Achieving End-to-End Segmentation

The challenge of semantic segmentation with scarce pixel-level annotations has induced many self-supervised works, however most of which essentially train an image encoder or a segmentation head that produces finer dense representations, and when performing segmentation inference they need to resort to supervised linear classifiers or traditional clustering. Segmentation by dataset-level clustering not only deviates the real-time and end-to-end inference practice, but also escalates the problem from segmenting per image to clustering all pixels at once, which results in downgraded performance. To remedy this issue, we propose a novel self-supervised semantic segmentation training and inferring paradigm where inferring is performed in an end-to-end manner. Specifically, based on our observations in probing dense representation by image-level self-supervised ViT, i.e. semantic inconsistency between patches and poor semantic quality in non-salient regions, we propose prototype-image alignment and global-local alignment with attention map constraint to train a tailored Transformer Decoder with learnable prototypes and utilize adaptive prototypes for segmentation inference per image. Extensive experiments under fully unsupervised semantic segmentation settings demonstrate the superior performance and the generalizability of our proposed method.