Airplane panels inspection via 3D point cloud segmentation on low-volume training dataset

This paper presents an approach for inspecting airplane fuselage panels using deep learning techniques, specifically segmentation models applied to 3D point clouds. A common challenge in deep learning applications is the limited availability of annotated training data — a problem particularly pronounced in industrial settings due to the high cost, time-consuming nature of data collection, and confidentiality concerns. Focusing on the detection and classification of rivets in the point cloud, our method aims to optimize the identification of potential defects, such as missing rivets or damaged rivets, even with minimal training data. Additionally, we evaluate the impact of different scanners and scanning resolutions on inspection results. This whole process serves as a preparation step in the reverse engineering pipeline.