Multi-scale Self-similarity Network for Few-Shot Segmentation

Few-shot segmentation is a specific type of segmentation in few-shot setting, using only a few number of annotated samples to segment unseen class objects in the given query samples. This challenging task faces the threats posed by large intra-class variations existing in support and query samples. To alleviate the inherent noise in matching relationship representation brought by intra-class variation, we present self-similarity representation by designing the multi-scale self-similarity module. It calculates the correspondence between the feature and its surrounding representation, which provides robustness for the network. Additionally, we further utilize the self-similarity representation by proposing the cross-attention decoder module to guide query feature in the decoding process. Comprehensive experiments on PASCAL-5i dataset in 1-shot and 5-shot paradigm demonstrate the superiority of our designed model. Our designed model significantly improves segmentation performance with only a few more expenses of training parameters.