标题：点云深度学习基准数据集

作者：杨必胜，韩旭，董震

来源：遥感学报

摘要：

为推进深度学习方法在点云配准、语义分割、实例分割等领域的发展,武汉大学联合国内外多家高等院校和研究机构发布了包含多类型场景的地面站点云配准基准数据集WHU-TLS和包含语义、实例的城市级车载点云基准数据集WHU-MLS。其中,WHU-TLS基准数据集涵盖了地铁站、高铁站、山地、公园、校园、住宅、河岸、文化遗产建筑、地下矿道、隧道等10种不同的环境,共包含115个测站、17.4亿个三维点以及点云之间的真实转换矩阵,为点云配准提供了迄今为止最大规模的基准数据集。WHU-MLS基准数据集涵盖了地面特征(机动车道、道路标线、井盖、非机动车道),动态目标(行人、车辆),植被(树木、树丛、低矮植被),杆状地物及其附属结构(电线杆、独立提示牌、路灯、信号灯、独立探头等),建筑和结构设施(房屋、道路隔离结构、围墙和栅栏)以及其他公共和便利设施(垃圾桶、邮筒、消防栓、街头座椅、电力线等)等6大类30余小类地物要素,共包含2亿多个点和超过5000个实例对象,为语义分割、实例分割点云深度学习网络的训练、测试和性能评估提供了当前最为丰富的基准数据集。

Abstract：

This paper aims to elaborate two large-scale point cloud benchmark datasets, namely, WHU-TLS and WHU-MLS, for deep learning purposes. The benchmark of the Whu-TLS data set comprises 115 scans and over 1740 million 3D points collected from 11 diff erent environments (i. e., subway station, high-speed railway platform, mountain, forest, park, campus, residence, riverbank, heritage building, underground excavation, and tunnel environments) with variations in the point density, clutter, and occlusion. The aims of the proposed benchmark are to facilitate better comparisons and provide insights into the strengths and weaknesses of di ff erent registration approaches based on a common standard. The ground-truth transformations and registration graphs are also provided to allow researchers to evaluate their registration solutions and for environmental modeling. In addition, the Whu-TLS data set provides suitable data for applications in safe railway operation, river surveys and regulation, forest structure assessment, cultural heritage conservation, landslide monitoring, and underground asset management. WHU-MLS benchmark dataset includes more than 30 kinds of objects and 5000 typical instances in urban scene. We manually labeled MLS point cloud, each point with spatial coordinates and normal. We totally labeled 40 scenes with average number of points 8 million, of which 30 scenes are split for training and 10 scenes for testing. The coarse and fine categories are defined as follows. The Construction: building (including the building façade and other clutters in the building), fence (including isolation structure on the road and wall); Natural: trees, low vegetation, including grass, shrub and other low tree; Ground: driveway (not including road mark), non-drive way, the ground that does not belong to the driveway, road markings; Dynamic: person (including person and bikes), car; Pole: light, electric pole, municipal pole, signal light, detector, board (usually attached to the light). The semantic labeling and instance labeling in WHU-MLS provide important references for point cloud deep learning. On the one hand, these datasets can be used for point cloud deep learning networks the training, testing, and evaluation of point cloud deep learning networks. On the other hand, the benchmark datasets would can promote the benchmarking of state-of-the-art algorithms in this field, and ensure better comparisons on a common base. WHU-TLS and WHU-MLS are freely available can be used freely for scientific research. We hope that the Whu-TLS and Whu-MLS benchmark data sets meet the needs of the research community and becomes important data sets for the development of cutting-edge TLS point cloud registration and point cloud segmentation methods.

表1 WHU-TLS基准数据集概述