Few-Shot 3D Point Cloud Semantic Segmentation via Stratified Class-Specific Attention Based Transformer Network

This paper presents a multi-layer transformer network for few-shot 3D point cloud semantic segmentation, addressing limitations in computational complexity and fine-grained relationship learning in existing methods. By aggregating query point cloud features with class-specific support features at multiple scales and avoiding pooling, our approach fully utilizes pixel-level support features.