Guided contrastive boundary learning for semantic segmentation

Semantic segmentation, a fundamental task in environmental understanding, aims to assign each image pixel to a specific class. Despite recent progress, segmentation accuracy in boundary regions remains suboptimal. This paper introduces Guided Contrastive Boundary Learning (GCBL), a novel framework designed to enhance feature representation learning, thereby improving boundary segmentation performance. Unlike conventional contrastive learning, GCBL guides inter-class representation learning by weighting pixel contributions based on their estimated probabilities. For intra-class learning, it leverages the neural collapse phenomenon, encouraging representations to align with last-layer classifier weights. Additionally, an asymmetric distance boundary pixel search strategy ensures a more reasonable selection of contrastive pairs. To prevent weight collapse in learning, a regularization term is applied to the last-layer classifier's weights. The GCBL method is readily integrable into existing and future segmentation frameworks. Extensive experiments on the Cityscapes, ADE20K, and S3DIS datasets demonstrate the effectiveness and generalizability of our approach. Code is available at https://github.com/skyshoumeng/GCBL.