SAM-MSF:A multi-scale indoor point cloud segmentation method based on deep learning

With the rapid development of 3D data processing technology, indoor point cloud segmentation has become a critical step for tasks such as scene understanding, target detection and path planning. However, existing methods are insufficient in dealing with complex geometric structures and multi-scale feature fusion. For this reason, this paper proposes a deep learning-based multi-scale indoor point cloud segmentation method (SAM-MSF). The method firstly utilizes the SAM model to segment RGB images at multiple scales and generates nested segmentation masks; subsequently, the 2D masks are mapped to 3D space with the help of the inverse projection method, and the local point clouds are integrated through a hierarchical and iterative merging strategy to obtain the global segmentation results.