Real-time Small Object Detection Model in the Bird-view UAV Imagery

Object detection is one of the most important parts of UAV applications. UAV imagery has object distortion and small-sized objects peculiarities. In this paper, we propose a D-RFB module which can enhance the expressive power of the feature map, and D-RFBNet300 attached D-RFB module so that detect small objects in the UAV imagery more accurately. And we propose the UAV-cars dataset including peculiarities of UAV imagery. Our D-RFBNet300 trained on MS COCO achieved 21% mAP with 45 FPS speed, which is the highest score among the other SSD type object detectors. In addition, our D-RFBNet300 trained on UAV-cars dataset achieved 99.24% AP at 10m altitude and highest AP at every test set altitude from 15m to 30m with 57FPS speed.