Hierarchical bidirectional aggregation with prior guided transformer for few-shot segmentation

Recent years have witnessed significant interest in few-shot segmentation methods, with the aim of predicting novel categories in a query image given the limited labeled support set. Despite demonstrated successes, some existing methods might suffer from the intra-class inconsistency between query and support samples for local unidirectional information guidance. We propose a hierarchical bidirectional aggregation with prior guided transformer for abundant intra-class common cues. Specifically, we adaptively aggregate support and query features by a non-local bidirectional information flow in a hierarchical manner to derive a closer and deeper correlation. We further introduce the prior affinity map to impart inductive bias and eliminate interfering semantics. Experimental results on three benchmark datasets demonstrate that the proposed method surpasses some previous state-of-the-art approaches well, especially performing favorably in handling challenging situations under 1-shot setting.