Bidirectional Prototype Learning With Differential Attention Denoising for Few‐Shot Point Cloud Segmentation

Abstract Semantic segmentation of 3D point clouds is constrained by the difficulty of obtaining labelled data. Few‐shot segmentation methods reduce the reliance on large‐scale annotations, but its performance is often constrained by the quality of prototypes, which can be degraded by noise and object feature discrepancies. To address this, we propose a Bidirectional Prototype Learning with Differential Attention Denoising (BPDA) framework. Our Differential Attention Denoising (DAD) module enhances prototype adaptation by suppressing irrelevant features and emphasizing critical ones. We further introduce a Bidirectional Prototype Learning mechanism, which iteratively refines prototypes by leveraging query predictions to generate new prototypes and inversely predict the class of support features. Furthermore, we incorporate a KL divergence loss to adjust the prototypes learning process through consistency constraints. Extensive experiments on S3DIS and ScanNet benchmark datasets demonstrate that our approach achieves competitive performance with state‐of‐the‐art methods.