DAUPNet: Domain-Aware Uncertainty Modeling for Reliable Prototype Discrimination in Cross-Domain Few-Shot Semantic Segmentation

DAUPNet addresses unreliable prototype matching in cross-domain few-shot semantic segmentation (CD-FSS) by combining hierarchical support-query feature harmonization, probabilistic foreground and background prototypes, and uncertainty-aware contrastive optimization. The paper reports average mIoU scores of 72.6% in the 1-shot setting and 76.7% in the 5-shot setting across four standard target domains, with particularly substantial gains on two medical domains. Its central claim is that prototype uncertainty can improve robustness and interpretability under severe domain shift. The authors also state that implementation code is available on GitHub.