Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Building-PCC: Building Point Cloud Completion Benchmarks

Weixiao Gao, Ravi Peters, Jantien Stoter

TL;DR

The paper tackles building point cloud completion in real-world urban LiDAR data by introducing Building-PCC, a benchmark pairing LoD2 building models with AHN3/AHN4 partial scans and ground-truth reconstructions. It evaluates eight representative deep-learning baselines using the mean Chamfer Distance $CD$-$l_1$ and $F$-$Score$ at $d=1\%$, revealing that real-world data introduce challenges not well captured by synthetic datasets. The study identifies key issues—data imbalance in local geometry, loss of fine details, and normalization hurdles—and discusses strategies such as synthetic data generation, multimodal inputs, and global-feature-aware networks to address them. This benchmark enables realistic assessment of building completion methods and guides future work toward more robust 3D geoinformation solutions, with code publicly available at the provided repository.

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 https://github.com/tudelft3d/Building-PCC-Building-Point-Cloud-Completion-Benchmarks.git.

Building-PCC: Building Point Cloud Completion Benchmarks

TL;DR

The paper tackles building point cloud completion in real-world urban LiDAR data by introducing Building-PCC, a benchmark pairing LoD2 building models with AHN3/AHN4 partial scans and ground-truth reconstructions. It evaluates eight representative deep-learning baselines using the mean Chamfer Distance - and - at , revealing that real-world data introduce challenges not well captured by synthetic datasets. The study identifies key issues—data imbalance in local geometry, loss of fine details, and normalization hurdles—and discusses strategies such as synthetic data generation, multimodal inputs, and global-feature-aware networks to address them. This benchmark enables realistic assessment of building completion methods and guides future work toward more robust 3D geoinformation solutions, with code publicly available at the provided repository.

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 https://github.com/tudelft3d/Building-PCC-Building-Point-Cloud-Completion-Benchmarks.git.
Paper Structure (17 sections, 4 equations, 7 figures, 1 table)

This paper contains 17 sections, 4 equations, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related work
  3. Point cloud completion datasets
  4. 3D shape completion
  5. The Building-PCC datasets
  6. Data collection
  7. Building instance extraction
  8. Benchmarks
  9. Implementation details
  10. Representative baselines
  11. Evaluation metrics
  12. Results
  13. Challenges
  14. Impact of imbalanced datasets
  15. Limitations on fine details
  16. ...and 2 more sections

Figures (7)

  • Figure 1: (a) and (b) display partial point clouds captured by airborne LiDAR, corresponding to AHN3 and AHN4, respectively. (c) presents a manually created LoD2 building model, while (d) illustrates point clouds sampled from (c).
  • Figure 2: FME workflow for extracting building point cloud instances.
  • Figure 3: Six visual examples of point cloud completion results by different approaches on the Building-PCC dataset. The color coding, ranging from blue to red, represents the height field, with blue indicating lower elevations and red signifying higher elevations.
  • Figure 4: AdaPoinTr prediction example using an AHN4 partial point cloud as input (refer to Figure \ref{['fig:fig1b']}). The area within the black circle highlights the roof parts that were not successfully predicted. The color coding from blue to red indicates elevation from low to high.
  • Figure 5: An example of details loss in the predict building point cloud using AdaPoinTr. The black circle emphasizes the absence of detailed roof superstructures.
  • ...and 2 more figures