Cycle association prototype network for few-shot semantic segmentation

Few-shot segmentation aims to train a segmentation model that can quickly adapt to novel classes referring to only a few annotated samples. Existing few-shot segmentation methods are based on the meta-learning strategy and extract support samples' information from a support set and then apply the information to make predictions on query images. However, most methods abstract support features into prototype vectors and ignore the crucial relationship between query and support samples. To address the problem, we propose a cycle association prototype network that focuses on pixel-wise relationships between support and query images for more accurate segmentation. Specifically, a cycle-consistent prototype module is proposed to select reliable support features and to generate prototype. To capture cross-scale relations and overcome object variations, we introduce a scale-aware prior mask generation module to offer rich guidance for objects of varying sizes and shapes via calculating the pixel-level similarity between the support and query image features. Finally, a mask generation module, which contains two parallel modules, feature fusion module and transformer decoder, is utilized to predict the query image. Extensive experiments on two datasets show that our method yields superior performance with state-of-the-art methods.