Coarse-to-Refined Category Reasoning and Propagation for Weakly Supervised Semantic Segmentation of 3D Point Clouds

Abstract

To mitigate the complexity of fully supervised point cloud semantic segmentation, research on weakly supervised methods has gained increasing attention. However, achieving semantic understanding with minimal annotations remains a significant challenge. Therefore, fully utilizing and extending the available annotations is essential. In this paper, we propose C2R-CR, a prototype-driven framework for Coarse-to-Refined Category Reasoning. First, a dynamic library of category prototypes is updated through point cloud annotation learning to facilitate category reasoning. Secondly, adaptive prototype enhancement leverages the dynamic prototype library to refine coarse category representations. Finally, the inferred category information is progressively propagated through hierarchical network layers to better guide model inference. Specifically, even with only 0.01% of annotated points, our approach achieves 66.2% mIoU on S3DIS Area 5. Our method provides valuable insights for other weakly supervised point cloud tasks, thereby contributing to the advancement of the field.

Keywords

Weakly Supervised / Point Cloud / Semantic Segmentation / Coarse to Refined / Category Reasoning

RIGHTS & PERMISSIONS

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.