Measuring Affinity: Similarity-Based Auxiliary Unlabeled Guidance for Few-Shot Segmentation

Yao Shen, Chunmeng Liu, Hanlin Chen, Kaiyang Zeng, Guangyao Li

Abstract: Due to the limited information derived from the support set, existing Few-Shot Segmentation (FSS) methods still suffer from the intra-class variation issue. To address this bottleneck, we propose a similarity-based Auxiliary Unlabeled Guidance Network (AUGNet), extracting auxiliary information from unlabeled images as a supplement of the original support guidance. First, the Auxiliary Information Generation module (AIG) refines unlabeled information through the pixel-wise corresponding similarity maps. Second, the attention-based Auxiliary Aggregation Module (AAM) integrates the auxiliary information with the original support guidance. To this end, a contrastive-learning based loss is designed to ensure the information consistency. Extensive experimental results with two different backbones on two challenging benchmarks demonstrate that our AUG, as a generic plugin, consistently achieves improvements over several leading methods under both 1-shot and 5-shot settings. Furthermore, we remove thoroughly the support set and apply the AUGNet under zero-shot setting. Experimental results illustrate its equal effectiveness under this extended segmentation task.