ScaleMatch: Multi-scale Consistency Enhancement for Semi-supervised Semantic Segmentation

• DOI: 10.1609/aaai.v39i6.32631
• Publication Date: 2025-04-11T11:12:03Z
• Authors: Liang Lv; Lefei Zhang

Semi-supervised learning improves semantic segmentation performance by leveraging unlabeled data, thereby significantly reducing labeling costs. Previous semi-supervised (S4) methods explored perturbations at the image level but neglected to adequately utilize multi-scale information. When labeled information is insufficient, scale variation between different objects makes instances with extreme scales even more difficult. To address this issue, we propose ScaleMatch, which aims learn scale-invariant features obtaining a mixed dual-scale pseudo-label and consistency learning. Specifically, cross-scale interaction fusion (CIF) module enforces interactive across scaled-views, allowing for reliable generation. More importantly, ScaleMatch introduces variable branches supervision. It consists of image-level (ISVC) feature-level (FSVC). Consequently, our enhances model's generalization under variation, outperforming existing state-of-the-art on both Pascal VOC Cityscapes datasets various partition protocols.