Dual Prototypical Contrastive Learning for Few-shot Semantic Segmentation

We address the problem of few-shot semantic segmentation (FSS), which aims to\nsegment novel class objects in a target image with a few annotated samples.\nThough recent advances have been made by incorporating prototype-based metric\nlearning, existing methods still show limited performance under extreme\nintra-class object variations and semantically similar inter-class objects due\nto their poor feature representation. To tackle this problem, we propose a dual\nprototypical contrastive learning approach tailored to the FSS task to capture\nthe representative semanticfeatures effectively. The main idea is to encourage\nthe prototypes more discriminative by increasing inter-class distance while\nreducing intra-class distance in prototype feature space. To this end, we first\npresent a class-specific contrastive loss with a dynamic prototype dictionary\nthat stores the class-aware prototypes during training, thus enabling the same\nclass prototypes similar and the different class prototypes to be dissimilar.\nFurthermore, we introduce a class-agnostic contrastive loss to enhance the\ngeneralization ability to unseen classes by compressing the feature\ndistribution of semantic class within each episode. We demonstrate that the\nproposed dual prototypical contrastive learning approach outperforms\nstate-of-the-art FSS methods on PASCAL-5i and COCO-20i datasets. The code is\navailable at:https://github.com/kwonjunn01/DPCL1.\n