Oriented-YOLOv5: A Real-time Oriented Detector Based on YOLOv5

Abstract

Current state-of-the-art oriented detectors have time-consuming feature extraction backbones, oriented proposals generation methods or additional special branches. These tricks increase the computational cost of oriented detectors and prevent them from some practical applications. YOLOv5 is one of the best real-time (inference time⩽33.3ms, or FPS⩾30) detectors in the field of general target detection that can be applied in various real tasks. However, YOLOv5 does not output the angular prediction that is crucial to reflect attitudes and shapes of the targets. We propose Oriented-YOLOv5 that can output angular prediction of rotated target and achieve real-time detection in aerial images. Specifically, we integrate Circular Smooth Label (CSL) into YOLOv5 (v5.0), so it inherits both the fast and lightweight features of YOLOv5 and the high performance of CSL for angle detection. Experimental results indicate that Oriented-YOLOv5s achieves an accuracy of 68.24% mAP and an efficiency of 11.8ms with aerial images of 1024×1024 input size. It can be further improved (68.86% mAP and 10.8ms) using Conv instead of Focus. We also integrate CSL into YOLOv5m and YOLOv5l, and they achieve improved accuracies (70.03% mAP and 71.2% mAP, respectively) and retain real-time inference speed. Overall, Oriented-YOLOv5 is well suited for detection tasks that have vertical view and require angular prediction with highly real-time and limited memory, so it can be used for applications deployed to satellites, UAVs, or robotic arms. Source code: https://github.com/Leesanjin/Oriented-YOLOv5