[论文解读] Large-scale 3D building and tree datasets constructed from airborne LiDAR point clouds in Glasgow, UK

LI, Qiaosi, Zhao, Qunshan+6 more|arXiv (Cornell University)|Jan 1, 2021

3D Shape Modeling and Analysis参考文献 1被引用 7

一句话总结

本文提出语义查询网络（SQN），一种仅使用0.1%随机标注点的弱监督3D语义分割方法，在仅使用0.1%随机标注点的情况下实现了最先进性能。通过基于查询的潜在表征机制利用局部邻域语义同质性，SQN 显著降低了标注成本，同时在 S3DIS、ScanNet 和 SemanticKITTI 等大规模数据集上优于完全监督基线模型，在 SemanticKITTI 上实现了55.9%的mIoU，参数量仅为1.05M。

ABSTRACT

IMPORTANT NOTE: The building data in this version (Version V1) is the legacy version. We suggest using the building footprint, lod1, and lod2 data of Version v4 (https://zenodo.org/records/15000747). The tree canopy and terrain data are the most updated in this version (Version V1). Urban Big Data Centre of the University of Glasgow generates 3D city models via the airborne LiDAR point clouds acquired between 2020-2021 on behalf of Glasgow City Council. It is a large-scale 3D city model containing 3D information on terrain, trees, and buildings in Glasgow City. This dataset comprises terrain, tree canopy, and building products derived from high-density airborne LiDAR point clouds. The terrain products include Digital Terrain Model (DTM), Digital Surface Model (DSM), and normalized Digital Surface Model (nDSM) in 0.5 m spatial resolution. The DTM and DSM rasters were provided by the vendor and nDSM rasters were obtained by subtracting DTM from DSM. Terrain products are provided in 5 km by 5 km GeoTIF format raster. The tree canopy products are composed of canopy height models (CHM) and tree top locations. Classified tree point clouds were applied with pit-free algorithm to generate CHM in 0.5 m grid raster in GeoTIF format [1]-[2]. Treetop locations were identified by using Local Maximum Filter based on CHM and are recorded as points in Shapefile format. The tree canopy products are provided in 5 km by 5 km tiles. Building 3D model products include footprint polygons with building height attributes and 3D mesh of building models in LoD1 and LoD2 levels. A series of processes such as converting building point clouds to building height models (BHM), converting BHM to polygons, and polygon regularization were conducted to obtain the building footprint polygons. Building height attributes were calculated from BHM for each footprint. The building footprint data are provided in Shapefile format. LoD1 models were generated based on the footprint and average height of the building. LoD2 models were constructed based on footprint and building point cloud with City3D tool[3]. LoD1 and LoD2 models are provided in OBJ and shapefile format. Building 3D model products are provided in 5 km by 5 km tiles. The RMSE of Euclidean distances between each point in the point cloud to the reconstructed model was calculated to evaluate the LoD2 model construction. A table of RMSE and a note for a few problematic models are provided.

研究动机与目标

• 研究在标注预算大幅减少的情况下，完全监督3D语义分割模型的性能极限。
• 解决在大规模3D点云中，特别是城市尺度场景下，完整标注成本过高且不切实际的问题。
• 开发一种弱监督方法，有效利用点云中局部语义同质性，以在极少监督下提升性能。
• 构建一种轻量化、端到端可训练的模型，避免复杂的多阶段流水线，适用于实际部署。

提出的方法

• SQN 使用预训练的特征提取器，将整个原始点云编码为分层潜在表征。
• 以任意3D点作为输入，查询其局部邻域的潜在表征子集。
• 通过可学习的聚合模块将查询到的表征汇总为紧凑向量。
• 该紧凑向量通过多层感知机（MLP）进行前向传播，预测最终的语义标签。
• 通过聚焦邻域结构，显式建模局部语义相似性，从而在稀疏标签下增强特征学习。
• 模型仅使用0.1%随机选择的点标注进行端到端训练，避免了复杂的数据增强或自监督预训练阶段。

实验结果

研究问题

• RQ1 当仅使用1%随机标注点时，现有完全监督3D语义分割模型的性能如何退化？
• RQ2 在何种标注比例（如0.1%）下性能显著下降，标志着弱监督阶段的开始？
• RQ3 在极端标签稀缺条件下，能否有效利用3D点云邻域中的局部语义同质性来提升分割性能？
• RQ4 在极少标注下，简单的端到端模型能否优于复杂的多阶段弱监督流水线？
• RQ5 是否可能仅使用0.1%标注点实现具有竞争力的性能，同时保持模型效率和可部署性？

主要发现

• SQN 在 SemanticKITTI 数据集上仅使用0.1%标注点即达到55.9%的mIoU，优于多个完全监督基线模型。
• 在 S3DIS 数据集上，SQN 使用0.1%标注训练时，分割精度与完全监督的 RandLA-Net 相当或更优。
• 在仅1.05百万个可训练参数下，SQN 在 SemanticKITTI 上实现50.8%的mIoU，展现出高效率和实际部署的适用性。
• 当标注比例低于1%时，性能迅速下降，表明0.1%是弱监督成为必要的临界阈值。
• 在相同0.1%标注预算下，SQN 在全部七个大规模基准数据集上均优于其他弱监督方法。
• 消融研究证实，局部邻域查询与语义同质性建模是该方法成功的关键，当移除这些组件时性能显著下降。