Data Augmentation for Deep Learning Training with Virtual Point Clouds Generated from CAD Models

In industrial plants, piping and equipment are intricately interconnected, and there are many components with a variety of shapes. To use point clouds of industrial plants for simulation of maintenance work, it is necessary to extract the components. Deep learning is effective for recognizing components in point clouds of industrial plants. However, training classifiers is challenging due to the difficulty in acquiring diverse point cloud datasets and the labor-intensive process of annotating large-scale point clouds. A promising approach to address these issues is to train the classifier on virtual point clouds generated from CAD models. However, classifiers trained on these virtual point clouds often fail to achieve sufficient segmentation accuracy due to discrepancies between virtual point clouds and actual point clouds captured by terrestrial laser scanners. This paper proposes methods to improve segmentation accuracy by reducing these discrepancies. First, we introduce a method to incorporate features such as missing points, noise, and outliers observed in actual point clouds. Furthermore, we propose a data augmentation approach that applies up-sampling using a deep learning model trained on paired virtual and real point clouds to reduce the discrepancy between them. Our evaluation demonstrates that the proposed methods effectively improve the segmentation accuracy of point clouds of industrial plants.