Domain Adaptation With Contrastive Learning for Object Detection in Satellite Imagery

Abstract

State-of-the-art (SOTA) object detection methods applied to satellite and drone imagery largely fail to identify cross-domain small and dense objects. The high content variability in the overhead imagery is due to different sensors, terrestrial regions, lighting conditions, and the image acquisition time of the day. Moreover, the number and size of objects in aerial imagery are very different than in the consumer data. We propose a small object detection pipeline that improves the feature extraction process by spatial pyramid pooling, cross-stage partial networks, and heatmap-based region proposal networks (RPNs). Next, we propose the instance-aware image difficulty score (DS) that adapts the overall focal loss to improve object localization and identification. Finally, we add the two progressive domain adaptation (DA) blocks using contrastive learning in the pipeline. The blocks align the local and global features extracted from the customized CSP Darknet backbone, as different levels of feature alignment alleviate the degradation of object identification in previously unseen datasets. We create a first-ever DA benchmark using contrastive learning for the object detection task in highly imbalanced satellite datasets with significant domain gaps and dominant small objects from existing satellite benchmarks—the proposed method results in up to a 7.4% and 4.6% increase in mean average precision (mAP) over the best SOTA method for the DOTA and NWPU-VHR10 datasets, respectively.