Deep learning-based 3D semantic segmentation: a practical case in hydrualic structures

Authors: Tianyang Chen*, University of North Carolina - Charlotte, Wenwu Tang, Associate professor
Topics: Geographic Information Science and Systems, Hazards, Risks, and Disasters
Keywords: Deep learning, point cloud, 3D GIS, hydrualic structure, semantic segmentation
Session Type: Virtual Paper
Presentation File: No File Uploaded

Deep learning-based classification algorithms in a 3D context have become a hot topic in recent years with an increase in accessibility of LiDAR data, where an increasing number of portable platforms (e.g., Drone, mobile LiDAR, static LiDAR) are able to collect 3D point cloud. Researchers have developed deep neural networks to accomplish semantic segmentation in different contexts, such as road marking detection, animal detection, and vehicle detection. In this study, we utilize a state-of-the-art 3D semantic segmentation framework, ConvPoint, in the context of hydraulic structure detection. Specifically, we aim to automatically detect hydraulic structures (e.g. bridges and their components, etc.) from the LiDAR scans in the fieldwork. However, rather than developing a general neural network relying on benchmark datasets, developing a deep neural network for a specific context (i.e. hydraulic structures in our case) can face challenges due to, for example, the availability and complexity of training data. We leverage transfer learning by adopting the related benchmark datasets to resolve this challenge. Moreover, we train the deep neural network on the data acquired by the LiDAR instrument during fieldwork and labeled manually. This case study can benefit the use of deep learning-based semantic segmentation in the context of hydraulic structures and provide empirical knowledge to studies in related contexts.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login