GeoKDS4:Geometry-Aware Distillation for Semi-Supervised Semantic Segmentation

Abstract

A key challenge in semi-supervised semantic segmentation (SSSS) is effectively leveraging unlabeled data to improve models trained with limited annotations. However, existing methods predominantly rely on visual cues, limiting models to the appearance-centric view of the data while overlooking geometric information, leading to suboptimal performance, particularly along object boundaries. Motivated by this insight, we propose a Geometry-aware Knowledge Distillation framework for Semi-Supervised Semantic Segmentation (GeoKDS4) that transfers geometric priors from geometry teacher model to a semantically focused student model. This process faces two challenges: (1) architectural differences between them lead to misaligned representations and (2) noise from the teacher’s insufficient geometric modeling in complex regions may disrupt the distillation process. To address these issues, we introduce two components: a Correlation-Guided Geometric Knowledge Distillation (CGKD) module to mitigate representation discrepancies by explicitly capturing feature correlations for accurate alignment and a Superpixel-driven Local Correction module to alleviate noise caused by weak local geometric representations by leveraging superpixel-based boundary priors to refine local regions. Extensive experiments on Pascal VOC and Cityscapes demonstrate that our method consistently outperforms state-of-the-art semi-supervised approaches across various labeled protocols, validating the effectiveness of our geometry-aware distillation strategy.