Deep feature extractors for small object detection in aerial images

Abstract

Convolutional Neural Networks (CNNs) are considerably developed year by year for better accuracy. New CNN architectures such as residual network (ResNet) are believed to replace the old ones (e.g. VGG) in all tasks. In this paper, we give evidence to prove that VGG is still useful for some specific tasks. We focus on training RetinaNet object detector on VisDrone dataset using ResNet and VGG as the backbone to detect four types of vehicle, which usually occupy small numbers of pixels in aerial images. From our experiments, we evaluate the effect of two mentioned CNN types on detecting small objects in images and undergo some incredible works from VGG – a sequential neural network.