Answers are grounded in this paper and cite their sources.
Dense pixel annotation constrains the scalable engineering deployment of road-scene perception systems, while conventional semi-supervised methods remain vulnerable to class-dependent confidence bias and the reinforcement of pseudo-label errors. We propose RSSSeg, a semi-supervised framework that couples class-aware reliable pseudo-label mutual learning with prototype-guided class-level contrastive learning. Two independently initialized segmentation branches supervise each other on unlabeled images. Class-specific confidence thresholds and prototype-to-pixel semantic consistency jointly select reliable pseudo-labels, after which cross-branch class prototypes provide compact positive targets and class-discriminative negative targets in the embedding space. RSSSeg achieves the highest numerical mean intersection over union (mIoU) across four labeled-data ratios on Cityscapes and CamVid, with average margins of 1.1 and 1.4 percentage points over RankMatch, respectively. Repeated-run experiments and component, pseudo-label-quality, sensitivity, efficiency, and adverse-condition analyses further assess the reliability mechanisms and practical behavior of the proposed framework. These results position RSSSeg as an annotation-efficient training framework for road-scene perception under limited supervision and adverse visual conditions.
No takes yet. Share an insight, caveat, or question.
Qu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: