Compact One-Stage Object Detection Network

Abstract

The targets in aerial images captured by drones are difficult to detect due to their small size, those neural networks with better detecting accuracy are too complicated to run real-time job on drone-mounted computer. This paper proposes a network combined residual network and YOLOv3-Tiny, residual network is used to merge different level features for improving YOLOv3-Tiny's small object detecting performance. During the experiment, the proposed network gets 2.9 higher mAP than YOLOv3-Tiny.