CAD4Real: Training Drone Object Detection Models without Manual Ingestion
Haixuan Wang, Lai Kang, Shanshan Wan, Tianrui Shen, Jingqian Gong · 2024
The performance of current pre-trained models for unmanned aerial vehicle (UAV) detection is directly impacted by dataset quality, and the manual ingestion and labeling required for dataset updates are both time and labor consuming. We present an automated procedure to constructing UAV synthetic datasets utilizing 3D photorealistic graphics. By implementing a targeted training strategy, we significantly enhance the performance of object detection algorithms under similar conditions. Experimental results demonstrate that, compared to a baseline training model, our model, trained with only 100 data samples, achieves a 64.9% increase in the [email protected] metric relative to the original model. Moreover, with 5,000 data samples, our model outperforms the performance achieved by the traditional model trained on 10,000 samples. This reduction in the requirement for real drone data samples streamlines the training process for UAV detection models.