Self-supervised deep learning methodology for automatic object detection in three-dimensional sensor systems

Abstract

This study introduces an innovative self-supervised deep learning methodology specifically designed to address the data scarcity problem in machine and deep learning frameworks, particularly for object detection in three-dimensional sensor systems. Using the enhanced data quality obtained from Fringe Projection Profilometry (FPP), this study showcases a pioneering approach that employs single-shot learning and advanced 3D data augmentation to enable precise multiscale object detection with minimal requirements for labeled training datasets. This work represents a pivotal contribution to the fields of 3D metrology and computer vision, offering a scalable and cost-effective solution that is particularly beneficial for precision-critical applications, such as manufacturing or medical imaging. The successful application of this methodology across a variety of experimental setups demonstrates its potential to significantly advance the practical utility and adoption of 3D sensing technologies in diverse sensor-based industries.