Few-Shot Segmentation via Complementary Prototype Learning and Cascaded Refinement.

Prototype learning has been widely explored for few-shot segmentation. Existing methods typically learn the prototype from the foreground features of all support images, which rarely consider the background similarities between the query images and the support images. This unbalanced prototype learning strategy limits its capability to mutually correct the segmentation errors between the foreground and background. In this paper, we propose a Complementary Prototype Learning and Cascaded Refinement (CPLCR) network for few-shot segmentation. Firstly, both the foreground and background features of the support images are used to learn our complementary prototypes. Then, the foreground and background similarity maps are jointly derived between the query image feature and our complementary prototypes, which capture more comprehensive prior information. Finally, we fuse the query image feature, foreground prototype and the foreground/background similarity maps together, and feed them to a cascaded refinement module, which recursively reuses the output of previous iteration to refine the segmentation result. Extensive experimental results show that the proposed CPLCR model outperforms many state-of-the-art methods for 1-shot and 5-shot segmentation.