Prototype-to-Prototype Matching for Few-Shot Semantic Segmentation
A few-shot segmentation strategy that compares learned support and query prototypes against each other, rather than matching a support prototype against raw query pixels. Prototypes are aggregated from Segment Anything Model encoder features through iterative cross-attention, whose attention weights double as a refined foreground pseudo-mask, and ResNet features are fused in for class consistency before the matched prototypes prompt the SAM mask decoder.
2501.00752
Few-Shot Segmentation (FSS) segments target regions in unlabeled query images given only a handful of labeled support images. This paper introduces Foreground-Covering Prototype Generation and Matchi…