Mining semantic information from intra-image and cross-image for few-shot segmentation

In recent years, few-shot segmentation has been proposed to alleviate the scarcity of pixel-wise labels, which performs segmentation on new categories using only a few annotated samples, while the problems of category-agnostic and low-data make few-shot segmentation very challenging. To address the task, we propose a new symmetric network, which mines semantic information from intra-image and cross-image in a holistic view and guides the segmentation of the paired images (i.e., the support image and the query image). We emphasize the importance of self-correlations in intra-image and inter-correlations in cross-image. Taking advantage of the provided labels, a self-attention relation module is proposed to transfer more category information for non-linear relation metrics by mining intra-image semantics. A co-attention module is designed to obtain common semantic information by exploring long-range dependencies of cross-image in spatial and channel dimensions, thus producing more precise segmentation results for the few-shot segmentation task. Experiments on two benchmark datasets (FSS-1000 and PASCAL-5i) show that the mean Intersection-over-Union scores of our method attain state-of-the-art performance.