Aerial Remote Sensing Object detection using Unsupervised Domain Adaptive

Abstract

Object recognition and localization in Aerial Remote Sensing Images (ARSI) are critical and demanding subjects for further processing object-related data in civil and military applications. To train a Deep Learning (DL) model for visual recognition and localization, a huge number of annotated images are needed. However, data categorization and annotation become a hard and time-consuming task. Despite the shortcoming of data in training, Unsupervised Domain Adaptation (UDA) offers an alternative solution to this issue. In this paper, UDA is suggested to detect and localize objects in ARSI as an unlabeled target domain. We compare the effectiveness of Faster Region Convolutional Neuronal Network (Faster R-CNN) as two stages detector and RetinaNet as one stage detector. These algorithms are based on the same Resnet50 model as the backbone. This study uses the natural image dataset MSCOCO as the source domain. We assess the proposed approach on two unlabeled datasets UC Merced and MTRASI datasets. The proposed method significantly improves object detection and localization performance, according to both qualitative and quantitative results. Extensive experiments show that the RetinaNet detector is better than the Faster R-CNN detector in terms of mAP.