Evaluating RTMDet for Efficient Small-Object Detection in UAV Remote Sensing Imagery

Unmanned aerial vehicles (UAVs) are capable of acquiring high-resolution remote sensing imagery in near real time, making them indispensable tools in emergency response, disaster relief, environmental monitoring, and public safety management. For the automated extraction of targets from such imagery, it is essential to systematically evaluate existing deep learning techniques to identify solutions that are best suited for practical deployment. This study investigates the performance of the Real-Time Models for object Detection (RTMDet) on the VisDrone2019-Det dataset, assessing five model variants—XL, L, M, S, and T—in terms of detection accuracy and computational efficiency, and benchmarking them against leading detection algorithms. The results show that: (1) RTMDet-XL achieves average precisions (AP) of 24.6% on the validation set and 20.5% on the test set, with the latter slightly underperforming compared to state-of-the-art models such as CornerNet (23.4%); (2) as the model is scaled down from XL to T, the number of parameters decreases from 94.9M to 4.8M, with corresponding drops in AP by 6.7% and 5.7%, respectively; and (3) RTMDet performs well in detecting large objects (e.g., vehicles), but continues to face challenges in identifying small targets (e.g., pedestrians). These findings highlight both the current capabilities and limitations of small-object detection in UAV imagery, offering valuable insights for further algorithmic refinement and practical application.