Testbed development for multisource dataset generation in machine learning object recognition

This paper presents a novel approach to synthetic data generation aimed at enhancing Machine Learning (ML) algorithms for object recognition. We detail the design and implementation of a testbed for capturing image data of 3D CAD models. This testbed facilitates the creation of datasets comprising images of 3D-printed scale models and rendered objects using the same CAD models, to be used for image recognition and associated ML tasks. The paper describes the development of an imaging system incorporating cameras, computer-controlled lighting and a 6-coordinate rotational/positioning system. In parallel, we generated 3D Blender renderings for comparison with both testbed-captured and real-world data. Generated data, similar in nature to aerial imagery targets, include synthetic images derived from CAD models and images captured from 3D-printed scale models. Tests involve various 3D models, image transformations, variations in the imaging process, and different training/testing data splits for image recognition tasks. Using these datasets, we investigate how ML algorithms handle the complex relationship between shape and physics in image formation. The results of the ML analysis and underlying motivations are presented in a separate paper.