RSTS-YOLOv5: An Improved Object Detector for Drone-Captured Images

Despite the tremendous progress in object detection in recent years, object detection on drone-captured images is still a great challenge because of the large number of small objects that appear densely and obscure each other in drone-captured images. In order to solve the problem of difficult object detection on drone-captured images, we propose a robust and efficient deep learning network RSTS-YOLOv5. We constructed Res swin transformer stage (RSTS) based on Swin-Transformer stage to extract global and contextual information and embedded it in YOLOv5x to explore the position of the transformer-based structure added in the detection network. In addition, we propose a multi-scale data augmentation for object detection on drone-captured images, which can enhance the robustness of the model for different scale objects without introducing additional computations. Experimental results show that our proposed RSTS-YOLOv5 achieves a mAP of 34.72% on the VisDrone test-dev subset and 34.84% on the validation-dev subset. Specifically, RSTS-YOLOv5 generalizes well on various drone-captured scenes, and is extremely competitive in object detection tasks on drone-captured images.