A method for generating a synthetic dataset of armored targets: randomized scene synthesis training

In this paper, we present a novel methodology, Randomized Scene Synthesis Training (RSST), for generating synthetic datasets of armored targets to train deep neural networks effectively. RSST extends domain randomization techniques to address the challenges of collecting and annotating real-world armored vehicle data, which is often constrained by security and resource limitations. The method involves creating non-photorealistic armored target data within a graphics engine, enabling the rapid generation of large, accurately labeled datasets at a lower cost. Experiments show that a deep learning model trained on this synthetic dataset performs well in detecting armored targets in real-world images, achieving a mean average precision (mAP) of 0.692. By leveraging synthetic data, the quality and diversity of object detection datasets can be significantly improved, leading to enhanced robustness and generalization capabilities. This approach not only facilitates the training of highly accurate models but also opens new possibilities for computer vision research, particularly in overcoming the limitations of real-world data collection.