Number-Adaptive Prototype Learning for 3D Point Cloud Semantic Segmentation

3D point cloud semantic segmentation is one of the fundamental tasks for scene understanding and has been widely used in metaverse applications. Many recent methods learn a single prototype (classifier weights) each class, classify points according to their nearest prototype. However, learning only class limits model’s ability describe high variance patterns within class. Instead this paper, we propose use an adaptive number prototypes dynamically different With powerful capability vision transformer, design Number-Adaptive Prototype Learning (NAPL) model segmentation. To train our NAPL model, simple yet effective dropout training strategy, which enables adaptively produce The experimental results on SemanticKITTI dataset demonstrate that method achieves 2.3% mIoU improvement over baseline based point-wise classification paradigm.