Centrala begrepp
Comprehensive evaluation of state-of-the-art deep learning methods for completing building point clouds in real-world urban environments, revealing key challenges and proposing solutions to advance the field of 3D geoinformation applications.
Sammanfattning
This paper introduces the Building-PCC dataset, a new real-world benchmark for evaluating deep learning-based point cloud completion methods on building data. The dataset consists of 50,000 building instances from two cities in the Netherlands, paired with corresponding partial point clouds from airborne LiDAR scans (AHN3 and AHN4) and complete ground truth point clouds sampled from manually reconstructed 3D building models.
The authors conduct a comprehensive evaluation of eight representative deep learning methods for point cloud completion, including PCN, FoldingNet, TopNet, GRNet, SnowflakeNet, PoinTr, AnchorFormer, and AdaPoinTr. The performance of these methods is assessed using the mean Chamfer Distance (CD-l1) and F-Score metrics.
The results show that PoinTr, AnchorFormer, and AdaPoinTr outperform other methods in terms of average CD-l1, with PoinTr achieving the best performance. AdaPoinTr excels in the F-Score metric, closely following PoinTr in CD-l1. However, the authors identify several key challenges faced by these methods when dealing with real-world building point clouds, including:
Imbalanced datasets: The uneven distribution of incomplete areas, such as building facades and roofs, can lead to models performing better on certain regions and worse on others.
Limitations on fine details: The methods struggle to accurately reproduce small building components, like chimneys and dormers, due to the low resolution of the predicted point clouds.
Over-smoothed sharp features: The methods tend to over-smooth sharp corners and edges of buildings, resulting in a loss of architectural details.
Normalization issues: Existing normalization techniques, which rely on aligning partial point clouds with ground truth data, may not be feasible in practical applications where ground truth is unavailable.
To address these challenges, the authors propose several potential solutions, such as developing synthetic datasets with diverse incomplete patterns, incorporating local plane distance into the loss function, and utilizing external data sources like GIS data to assist with normalization. These insights aim to guide future research and foster innovation in the field of 3D geoinformation applications.
Statistik
The mean Chamfer Distance (CD-l1) for the AHN3 and AHN4 datasets are:
PCN: 6.09 and 6.14
FoldingNet: 4.20 and 5.63
TopNet: 6.42 and 6.44
GRNet: 5.21 and 4.89
SnowflakeNet: 6.61 and 6.60
PoinTr: 1.40 and 1.40
AnchorFormer: 1.46 and 1.46
AdaPoinTr: 1.42 and 1.41
The F-Score for the AHN3 and AHN4 datasets are:
PCN: 0.374 and 0.416
FoldingNet: 0.328 and 0.352
TopNet: 0.262 and 0.267
GRNet: 0.399 and 0.463
SnowflakeNet: 0.588 and 0.629
PoinTr: 0.511 and 0.585
AnchorFormer: 0.631 and 0.685
AdaPoinTr: 0.679 and 0.725
Citat
"Despite deep learning methods excelling in capturing local and global geometric features, the quality of results in real-world scenarios is poor and fails to meet downstream applications' needs."
"Existing normalization techniques, which rely on aligning partial point clouds with ground truth data, may not be feasible in practical applications where ground truth is unavailable."