Bridging the Synthetic-to-Real Gap (BSRG): Creating Simulated Datasets for Domain Adaptation to Enhance Vehicle Detection

Deep neural network based military vehicle detectors pose particular challenges due to the scarcity of relevant images and limited access to vehicles in this domain. Moreover, Real-world data often poses significant challenges, including privacy, availability, and bias. To mitigate these challenges, synthetic datasets can be leveraged. This article introduces a synthetic dataset related to the Joint Light Tactical Vehicle (JLTV) aimed at object detection and examines this dataset in the context of training advanced object detection models. Using the powerful Unreal Engine, which can Produce extremely lifelike environments, we generated a comprehensive synthetic dataset designed to simulate real-world conditions and enhance the training process for various detection algorithms. In this study, we evaluate two distinct domain adaptation models for object detection: an enhanced domain matching approach utilizing the Masked Image Consistency (MIC) framework and an unsupervised domain matching approach employing confidence-based mixing (ConfMix). The MIC model achieved a mean Average Precision mAP@50 of 62.6% on real-world data, while the ConfMix model attained a mAP@50 of 55.8%. These results underscore the pivotal role of synthetic data in advancing object recognition technologies. They also highlight potential research directions for improving synthetic dataset generation and enhancing model performance in practical applications. Examples of this dataset can be accessed at: https://github.com/behnaz-sadeghigol/JLTV_dataset.