Building-PCC: Building Point Cloud Completion Benchmarks

Weixiao Gao,Ravi Peters,Jantien Stoter

2024-04-24

Abstract:With the rapid advancement of 3D sensing technologies, obtaining 3D shape information of objects has become increasingly convenient. Lidar technology, with its capability to accurately capture the 3D information of objects at long distances, has been widely applied in the collection of 3D data in urban scenes. However, the collected point cloud data often exhibit incompleteness due to factors such as occlusion, signal absorption, and specular reflection. This paper explores the application of point cloud completion technologies in processing these incomplete data and establishes a new real-world benchmark Building-PCC dataset, to evaluate the performance of existing deep learning methods in the task of urban building point cloud completion. Through a comprehensive evaluation of different methods, we analyze the key challenges faced in building point cloud completion, aiming to promote innovation in the field of 3D geoinformation applications. Our source code is available at

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper attempts to address the issue of incomplete urban building point cloud data collected through LiDAR technology in the real world, which often occurs due to factors such as occlusion, signal absorption, and specular reflection. These issues severely affect the quality of 3D reconstruction. Therefore, the paper aims to introduce a new benchmark dataset called Building-PCC to evaluate the performance of existing deep learning methods in completing urban building point cloud tasks and to analyze the key challenges to promote innovation in the field of 3D geographic information applications.

Specifically, the main objectives of the paper include:
1. **Constructing a new benchmark dataset**: Creating a Building-PCC dataset based on real scenes, containing point cloud data extracted from 50,000 building instances in the cities of The Hague and Rotterdam in the Netherlands.
2. **Evaluating existing methods**: Conducting a comprehensive evaluation of existing deep learning methods, particularly their performance in handling real-world building point cloud data.
3. **Analyzing key challenges**: Exploring the main issues encountered in the task of completing building point clouds, such as data imbalance and loss of structural details, and proposing potential solutions.

Through these objectives, the paper hopes to drive technological innovation in the field of 3D geographic information applications, improving the accuracy and reliability of point cloud data in practical applications.