Machine learning models have shown remarkable success in computer vision tasks, making them a promising tool for quality verification in manufacturing processes. However, these models require large volumes of annotated data for training, which may be impossible to obtain for new or low-volume manufacturing processes. In this study, we investigated methods to reduce the need for annotated data to enable detection of low probability of occurrence events. Our approach involved the generation of synthetic data using a 3D game engine to simulate variability in a part and training a computer vision model to detect defects in real images of the simulated part. We used a DB9 connector as our simulated part and classified a “bent pin” as our defect of interest. We utilized Unity Perception to generate annotations for our synthetic data. We evaluated the performance of our model on a small set of human-annotated images of a single DB9 connector with various pin bends. Our findings demonstrate the potential use of synthetic data to train computer vision models and also outline best practices when generating synthetic training data.