Evaluating CAD Usage to Generate Synthetic Images for Machine Learning-Based Object Detection

When implementing machine learning (ML) methods, for example, in the context of quality control in production, a major concern for small and medium-sized enterprises (SMEs) is the effort, time, and thus cost required to train a powerful ML model. Since nowadays most manufacturers already have computer-aided design (CAD) models of their products, using these for training an ML algorithm could be beneficial. In this paper, the performance of the YOLO Object Detection Algorithm is evaluated on images directly rendered from a CAD model in Autodesk Inventor and real-world images of the same product. The main goal is to reduce the number of real-world images needed to train the model using CAD-Renderings. This is especially useful before the start of production since real-world images are not available at that point. This will continue to enable faster adaptation to new products and thus drive digitization in companies.