Dynamic routing towards few-shot point cloud semantic segmentation

Current 3D point-cloud semantic segmentation employs few-shot learning to lessen reliance on large-scale data. Previous prototype-based methods typically extract and refine prototypes from support and query sets, which are followed by the cosine similarity or label propagation for semantic label prediction. However, these models often neglect global dependencies, resulting in the overlook of the prototype-level context between the query and support sets, which reduces the accuracy and robustness of few-shot point cloud semantic segmentation. To tackle this challenge, we propose an Adaptive-correlated Prototype-oriented dynamic Routing (APR) framework that employs dynamic routing to delve into the context of query-support prototypes, which helps to generate high-quality class prototypes for further reasoning. Furthermore, we utilize dynamic routing coefficients to explore the correlations between support and query prototypes, which are embedded in the loss function to bolster the accuracy of final label predictions. Extensive experiments conducted on the public S3DIS and ScanNet datasets highlight the superior performance of our proposed framework.