LSDX: Point Cloud Segmentation with Local Surface Distribution and X-Net

Abstract Purpose and Methods: Currently, most network models for point cloud segmentation face challenges such as inadequate capture of continuous structural information on the point cloud surface and low feature recognition rates. This paper first investigates the local surface structure of point clouds and introduces the Local Surface Distribution (LSD) structure for analyzing and learning point clouds. This structure calculates the distribution of local neighborhoods in a per-point manner, enhancing the continuity of structural information within the point cloud. Then we propose X-Net, a model structure based on continuous encoding and decoding. Based on LSD and X-Net, we present the LSDX model framework for point cloud segmentation. Furthermore, we specify two implementations of LSDX, namely LSDX-PN2 and LSDX-PT. Results and Conclusion: We evaluate our proposed LSDX on several challenging benchmarks, and experimental results demonstrate that LSDX produces more accurate segmentation boundary and significantly reduces isolated points in the prediction results than other methods. In particular, LSDX-PT, which takes Point Transformer as the baseline, outperforms the latest model Geospark and PointMetaBase on ScanNetv2 and S3DIS benchmarks, achieving state-of-the-art performance.