Terravehicle: A Million-Point-Per-Vehicle Dataset for Fine-Grained Component Segmentation

Abstract

Current vehicle point cloud datasets are challenged by incomplete geometric details, low annotation accuracy, and limited model diversity, which are difficult to apply to high-precision component segmentation tasks. So, a centimeter-accurate million-point-per-vehicle point cloud dataset for fine-grained component segmentation is developed, which is named TerraVehicle and is captured by six high-precision LiDARs. TerraVehicle employs a novel annotation strategy based on topology-constrained segmentation primitives, surpassing traditional 3D bounding boxes, and achieves 10x greater precision compared to KITTI, nuScenes, and A* 3D. The benchmark tests against ShapeNetPart show that classification-level mIoU scores for PointNet++, Point-Mamba, Point-PlaneNet, and Point2Vec on TerraVehicle remained competitive, with some models exhibiting performance improvements. Furthermore, instance-level mIoU scores for PointNet++, Point-PlaneNet, and Point2Vec on TerraVehicle outperform those on ShapeNetPart, which highlights that the proposed TerraVehicle  is very suitable for high-precision segmentation tasks. The experimental results validate the proposed TerraVehicle as a high-quality benchmark for advancing the development of point cloud component segmentation algorithms.

Keywords: 3D Point Cloud, vehicle measurement, component segmentation, Dataset, centimeter-accurate