Exploring State Space Models in LiDAR Point Cloud Segmentation

Abstract. Mamba has achieved significant success in various fields due to its ability to efficiently model long-range dependencies with linear complexity. However, its application in LiDAR point cloud processing is still in its early stages, facing challenges such as unordered and irregular data structures. In this study, we investigated the performance of two existing Mamba-based algorithms, PointMamba and PointCloudMamba, on the aerial DALES LiDAR dataset for point cloud segmentation, and further explored the critical role of token serialization in influencing Mamba’s performance. To evaluate serialization quality, we proposed two novel indicators—Neighbor Preservation Ratio (NPR) and Sequence Jump Distance (SJD)—which quantify the ability of serialization methods to preserve spatial topology and geometric relationships. Our findings confirm the great potential of Mamba in LiDAR point cloud processing, and demonstrate that serialization significantly impacts Mamba’s performance, with better preservation of spatial and geometric relationships leading to higher segmentation accuracy. These results provide meaningful insights into improving Mamba’s performance in LiDAR point cloud processing and guiding the development of advanced serialization methods.