UavTinyDet: Tiny Object Detection in UAV Scenes

Abstract

Due to the limited battery life, object detection using drones is often carried out with a large field of view and high flight altitude. In this case, the objects of UAV images tend to be tiny in size and are therefore not easy to be detected, which remains an open challenge. Although feature pyramid networks can improve this situation, it is still ineffective for predicting extremely small objects considering that information about tiny objects may have already vanished in feature maps. Moreover, current label assignment strategies of CNN-based detectors are usually not suitable for tiny objects due to their tiny size. In this paper, we propose a new method called Expanded Feature Pyramid (EFP) to take full advantage of the low-level information of the feature pyramid, which can better represent tiny objects with less information lost. Based on YOLOv5, to mitigate the deterioration of its performance by introducing many low-quality positive samples for tiny objects, we propose a novel prediction-aware loss for classification and localization, allowing the detector to select more high-quality positive samples which have higher IoU with ground truth objects and focus on training on them. Experiments on the VisDrone-DET dataset demonstrate our method achieves competitive performance among various object detectors. In addition, for the enrichment of VisDrone-DET, we construct a UAV dataset named UAV-OUC-DET with larger image size and abundant small objects and tiny objects. Experiments on the UAV-OUC-DET also demonstrate the effectiveness of our method.