Harnessing Vision Transformers for LiDAR Point Cloud Segmentation.

Point cloud data, representing 3D objects, has become an indispensable format in numerous applications. However, directly processing this data form, particularly for tasks like segmentation, remains a challenging endeavor. In this manuscript, we present an innovative methodology tailored for point cloud segmentation. Our approach exclusively harnesses the capabilities of two Vision Transformers (ViTs). To facilitate this, we first transform the point cloud into a 2D image representation through spherical projection. Subsequently, the initial ViT is utilized for the extraction of salient features from the projected image. These extracted features are then channeled into the second ViT. Our methodology synergizes the compactness and efficiency of spherical projection with the robustness and adaptability of vision transformer networks. The latter has been celebrated for its state-of-the-art performance across a myriad of computer vision challenges. For empirical validation, we train our network architecture using the renowned SemanticKITTI dataset, aiming to benchmark its efficacy against existing frameworks.