Generating Synthetic Image Data with Blender to Address Data Scarcity in Military Applications: Leveraging the RF-DETR Model

Abstract. Military vehicle recognition faces critical data scarcity due to operational security constraints and prohibitive collection costs. Classification of vehicles demands extensive training data rarely available in defence contexts. We propose a hybrid approach combining limited real-world data with scalable synthetic generation. Our methodology comprises: (1) a Blender-based pipeline generating high-resolution synthetic images with domain randomization across 3D models, lighting, and camera angles; (2) training transformer-based RF-DETR detectors on real-world and synthetic data, respectively; (3) an in-depth evaluation of the trained networks to determine the effect of synthetic data. Our approach utilizes a baseline RF-DETR detector trained on real-world imagery to compare against. Then we utilize the custom-made synthetic data generation pipeline to create an equally large synthetic dataset. This generated data is added to real data subsets, thus creating a mixed datasets containing varying percentages of real data. We created five datasets containing 5%, 10%, 25%, 50%, and 100%, respectively. With these new mixed datasets we train another set of RF-DETR detectors. Afterwards we evaluate the influence of the synthetic data by comparing the detectors across computer vision metrics.