Small Object Detection using Prediction head and Attention

UAV(Unmanned aerial vehicle)-captured image contains a number of small object. Object captured in UAV's low altitude flight are expressed as low resolution in the image and have ambiguous boundaries. The detection problem of small objects expressed in limited pixels in UAV-captured images is difficult. In this paper, we used an additional prediction head to improve the detection performance o small objects, and modified the channel attention module of CBAM and added it to the PANet. The experiment showed that the proposed method showed a 4.1% improvement in mAP performance compared to the existing method in VisDrone 2020-DET dataset (Abstract)