Comparative Analysis of Synthetic Data Generation for Object Detection: CAD Models vs. 3D Scans of Industrial Items and Hybrid Approaches

Deep learning techniques, particularly in object detection, are becoming increasingly common in industrial settings. However, the challenges posed by industrial objects such as intricate surface textures and complex geometries often require the creation of custom training datasets. Publicly available datasets typically do not provide sufficient coverage for these unique characteristics. In many cases, producing real-world datasets for high-mix, low-volume (HMLV) production scenarios is both time-consuming and costly. In this paper, we evaluate the use of synthetic data generated using NVIDIA's Isaac-Sim as an efficient alternative. We compare the use of CAD models and 3D scans of real assets, reconstructed using state-of-theart scene reconstruction methods, e.g. structured light scans and Neural Radiance Fields (NeRFs). For this, we utilize pre-existing hardware and software tools and set the focus on the usability in an Industry 4.0 environment. The generated synthetic datasets are used to train a YOLO-based object detection model for a worker assistance system that provides context-based assembly instructions. The model is tested with real image data of two objects with distinct surface and texture properties. Initial results demonstrate performance that exceeded expectations.