RangeSAM: Promptable Segmentation of Sparse 3D Point Clouds in Outdoor Driving Scenes

Abstract

Point cloud segmentation is critical for 3D scene understanding in autonomous driving. While promptable segmentation has emerged as a promising approach for flexible, user-guided object segmentation, existing 3D methods target dense point clouds and fail to handle the sparsity inherent in real-world LiDAR data from outdoor driving scenes. We introduce RangeSAM, a novel architecture for promptable segmentation of sparse point clouds that integrates range-based representations with a bi-directional transformer to address sparsity in LiDAR data. We further enhance this design with three targeted innovations: a learnable scaling factor for dynamically balancing prompt and image embeddings, a prompt loss function encouraging localized attention around prompts, and prompt label enforcement ensuring prediction consistency. Our model design enables effective segmentation through intuitive point-based prompts while capturing long-range dependencies critical for sparse 3D data. Evaluated on KITTI360, SemanticKITTI, and ApolloScape, RangeSAM achieves 56.05% IoU with 20 point prompts on KITTI360, demonstrates strong generalization with over 46% IoU on ApolloScape despite significant domain differences, and attains competitive performance against non-promptable methods on SemanticKITTI with an F1 score of 74.4 and segmentation association score of 50.4. Comprehensive ablation studies validate our architectural innovations, with their combination yielding a 7.97% improvement over the baseline. These results establish RangeSAM as an effective solution for adaptive segmentation of sparse point clouds in real-world applications.