Synthetic-to-real domain adaptation for UAV based object detections in trench environment
Phani R. Yalamanchili, Zhangyu Jin, Andrew Feng, Grant Spellman, Michal Harari, James Uplinger, Celso M. de Melo · 2025
Object detection is an essential task in various vision applications such as automatic target recognition, search-and-rescue, surveillance systems, and autonomous driving. Detecting objects such as moving persons from UAV-based aerial view is particularly difficult due to the smaller object sizes from higher altitudes. In this research, we focus on detecting persons in a trench environment, which adds further challenges due to significant occlusions when individuals are hiding within covers. Our method leverages synthetic data generation techniques coupled with domain adaptation strategies to address the challenge of data scarcity and improve model generalization. Recent advancements in synthetic-to-real object detections have highlighted the importance of addressing domain shifts, where models trained on one dataset may not perform well when deployed in a different environment. We address this problem through the creation of a synthetic dataset mimicking the camera views and environmental effects from the real dataset collected in open video repositories such as YouTube. We further apply image style transfer and domain adaptation techniques to enhance the model for a real-world benchmark. Our results show that our model outperforms models trained with only real data. We offer insight on solutions to optimize synthetic data and model training for this important use case.