MGNet: Mutual-guidance network for few-shot semantic segmentation

Few-shot semantic segmentation has recently drawn attention for its remarkable potential to segment the regions of different object classes with only a few labeled samples as guidance. Although recent methods have achieved impressive performance, there exist two critical bottleneck problems to be solved. First, most existing methods typically model a target class only using information from the foreground regions of support images, which actually does not adequately exploit the background region information of support images. Second, segmentation performance will be greatly affected when there is a large intra-class variation between the support and query images of the same class. To address these problems, we propose a mutual-guidance network (MGNet) for few-shot semantic segmentation to enhance the discriminative ability of class-specific prototypes. More specifically, the prototype learning module is first devised to learn the class-specific prototype of the foreground and background regions. Then, with non-parametric metric learning, the deep features of the query image are matched with multiple learned prototypes. Finally, to make good use of the ground truth mask of the support image, a reverse auxiliary learning module is constructed to reinforce the learned prototype. Extensive experiments on two standard benchmarks PASCAL-5(??) and COCO-20(??) are shown that the proposed method can yield competitive segmentation results with state-of-the-art methods. Surprisingly, our model achieves state-of-the-art results under both 1-shot and 5-shot tasks on more challenging COCO-20(??) when ResNet-101 is used as the backbone network.