Self-positioned Point Feature Retrospective Transformer: An Efficient and Fine-Grained Framework for Point Cloud Modeling

Transformers have achieved remarkable success in computer vision tasks. However, directly applying them to point cloud processing remains challenging due to their high computational complexity and the lack of dedicated mechanisms for modeling 3D data. We introduce the Self-positioned Point Feature Retrospective Transformer (SPFR) for efficient and accurate point cloud analysis. SPFR addresses challenges of computational complexity and limited 3D modeling by combining adaptive representative points with local self-attention and global cross-attention. Additionally, it incorporates a retrospective semantic refinement module that propagates semantic context across decoder layers through cross-attention and a semantic gating unit. This design enhances point-wise feature precision and consistency. Evaluations on ModelNet40, ShapeNet, and S3DIS datasets demonstrate that SPFR achieves excellent performance across various point cloud tasks. Qualitative results further validate the effectiveness and interpretability of the selfpositioning and refinement mechanisms. SPFR is broadly applicable to tasks such as classification, part segmentation, and scene segmentation in point cloud processing.