Investigating Deep Learning Colorization of Terrestrial Laser Scanning Point Clouds for Calibration-light and Rapid Visualization

Abstract

Terrestrial laser scanning (TLS) is widely used for generating accurate, high-resolution colored point clouds. However, a significant portion of the time at each scanning station is dedicated to capturing and stitching panoramic images for accurate colorization. For applications where precise colors are not critical, pseudo-colored point clouds can serve as a faster and computationally efficient alternative, particularly for presentations and qualitative analysis. This study proposes a novel pipeline leveraging advancements in deep learning-based photo colorization to pseudo-color point clouds. By treating the intensity maps from TLS as single-channel grayscale images, we apply pretrained deep learning models from the DeOldify library, trained with NoGAN techniques, to generate realistic pseudo colors. Indoor and outdoor TLS scan slices are showcased before and after pseudo colorization. The resulting, colored intensity maps are used to colorize point clouds, creating visually compelling representations. We compare these pseudo-colored point clouds with traditionally colorized counterparts that utilize RGB imagery. The results demonstrate that pseudo-colored point clouds can effectively enhance human evaluation and visualization tasks, providing a valuable tool for scenarios where real colors are unnecessary.