Evaluating the influence of backbone network architectures for object detection in aerial images

Abstract

Drones are increasingly being used in surveillance, agriculture, and delivery tasks. However, the real-life application of images collected from drones in urban manage- ment in Vietnam is still limited. Although drone images have many advantages thanks to the flexibility of the latest devices, there are still new challenges, such as top-down views, small objects, arbitrary directions, and class imbalance. In this paper, we conduct research, survey, and evaluate the performance of CNN-based network architectures on object detection in aerial images. Experiments were conducted on seven deep learning network architectures: VGG, ResNet, ResNext, Res2Net, ResNeSt, HRNet, and RegNet to bring objective judgments and conclusions based on experiments, contributing to the development of solutions for applications of determining the status of urban traffic in Vietnam.