.png){kind=icon}
.png){kind=icon}
Return
Uncertainty-aware point cloud segmentation for infrastructure projects using Bayesian deep learning
.png){kind=icon}
OA
AI
DOI：10.1016/j.autcon.2024.105419.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reliable traffic infrastructure is a key factor for any country's economy. However, aging bridges often require renovation or reconstruction. Promising approaches for enhancing asset management and cost reduction are digital twins and predictive maintenance strategies. However, the creation of geometric-semantic as-is models as a basis for digital twins currently involves labor-intensive manual data capture and modeling. Modern deep learning models, such as Kernel Point Convolution (KPConv) show promising results in reducing the time needed to create digital twins by semantically segmenting point cloud data but have so far been hindered by the lack of a reliable quality measure, which can predict when the model's prediction can be trusted. This is especially viable in the construction industry, where objects and sensors may vary widely between projects and contractors. In this work, we present Bayesian neural networks, implemented through Variational Inference and MonteCarlo dropout as approaches to conducting inference with KPConv, which show improvements in the confidence estimation and out-of-domain (OOD) detection in the scans of typical infrastructure point clouds. When confronted with different domain shifts to the test data, such as a change of scanning device and introduction of unseen classes the proposed model showed a 25.3% decrease in expected calibration error (ECE) and a 4.82 increase in OOD detection in terms of outlier-intersection-over-union (O-IoU) on average with respect to a deterministic baseline.
Keywords:
Uncertainty estimation
Semantic segmentation
Point cloud
BIM
Digital twin
Bridge
Bayesian deep learning
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
11.5
Papers:
6.2K
Citations:
4.2W