PMDAv2: Multi-scale prototype matching for domain adaptive semantic segmentation

Weiwei Li, Yuchen Zheng, Yuanyuan Ren, Junzhuo Liu, Yahao Liu, Wen Li

Abstract: highlights•We propose a novel approach named Prototype Matching Domain Adaptation v2 (PMDAv2) to address optimization deviation of distance-regularizing based domain alignment method.•We design a multi-scale prototype (MS) strategy to overcome the issue that falsely aligns different classes that share local appearances across domains by incorporating multi-scale spatial information into the prototypes.•We propose a confidence-based re-weighting prototype calculation method to mitigate the impact of noisy and ambiguous samples during prototype-based domain alignment.•We have conducted extensive experiments on three widely used benchmarks, and the experimental results show the effectiveness of our proposed method.

External IDs:doi:10.1016/j.patcog.2025.112771