Textureless CAD-based domain randomized synthetic data generation for AI model training

Training deep neural networks for object detection typically requires large-scale, annotated real-world datasets, which are costly and time-consuming to collect. To address this limitation, we propose a domain-randomized synthetic data generation pipeline that converts textureless CAD models into labeled synthetic datasets for training computer vision models. Specifically, we employ three distinct texturing strategies as input configurations to generate synthetic datasets: (1) using the original textureless CAD models, (2) applying randomized color augmentations to introduce appearance diversity, and (3) enhancing CAD models with realistic, physically based materials to simulate plausible surface properties. Each strategy is used to generate synthetic datasets for training YOLOv8 object detection models, and their effectiveness is systematically evaluated. Experimental results show that models trained on realistically textured synthetic data outperform those trained on textureless or randomly colored data, achieving a mean average precision of 0.909 on a real-world test set. Furthermore, by applying a one-shot transfer learning approach using only 1% of real data, we increase the mean average precision to 0.954, effectively narrowing the gap between synthetic and real-world performance. These findings highlight the potential of combining high-fidelity synthetic data with minimal real-world supervision for efficient and scalable object detection training in industrial applications.