Empirical knowledge related to deep learning-driven 3D point cloud classification in 3D GIS

Keywords: 3D GIS, deep learning, semantic segmentation, transportation, hydrology
Session Type: Virtual Paper Abstract
Day: Saturday
Session Start / End Time: 2/26/2022 02:00 PM (Eastern Time (US & Canada)) - 2/26/2022 03:20 PM (Eastern Time (US & Canada))
Room: Virtual 3

Abstract

In transportation, LiDAR is used to monitor the as-built status of the structures and track the geometry changes for safety management. Manually interpreting LiDAR data is time-consuming and labor-intensive to label the point cloud for subsequent processing. Semantic segmentation algorithms are designed to automate this process. Deep learning-based 3D classification algorithms are developed in recent years, and they appear to represent the state-of-the-art-performance. Deep learning-based methods require a quite large training dataset to achieve a reasonable performance. However, specialists may face limited availability of 3D data, which may further lead to imbalanced labeled classes (also called “long-tail”) distracting the model generalization, especially in a 3D content of a specific application, represented by hydraulic structures in this study. The development of mobile LiDAR increased its mobility and efficiency to collect 3D point clouds, spurring the emergence of related real-time applications in the 3D Geographic Information System (GIS). 3D GIS is featured by greater details with adding one more dimension. Capabilities of 3D GIS rather than visualization are still in early stages, such as GIS Augmented Reality, disaster response, safety management. We shared our empirical knowledge generated from a related NCDOT project, DeepHyd. Typically, we will share knowledge used in the development process, from training data preparation to hyper-parameter tuning. This will benefit the future practice of related applications using deep learning-based methods in 3D context. At last, we will discuss how deep learning-based methods can contribute to the data collection in 3D GIS especially for those mentioned applications in early stages.