Vehicle detection in UAV traffic videos using GAN online augmentation: a transfer learning approach

Low resolution object detection could be challenging. In this paper, we proposed a GAN-based real-time data augmentation algorithm for the transfer learning task of UAV vehicle detection from ImageNet, with improvements including using FocalLoss to replace the cross-entropy loss commonly used in the industry, as well as redesigning the target detection Head combination to improve the model's detection accuracy by 4% over original YOLOv5 model. We make it feasible for deployments on UAV-carried ARM systems.