Deep Domain Adaptation for Single-Shot Vehicle Detector in Satellite Images

摘要

In this paper, we designed unsupervised domain adaptation (DA) methods to vehicle detection in high-resolution satellite images. We applied two Single Shot MultiBox Detectors, which have advantages in handling image feature differences among various kinds of image data: Correlation Alignment DA (CORAL DA) and adversarial DA. These novel methods can much improve accuracy without annotated data by finding the common feature space of source and target domains and aligning the features. While a mean of average precision (AP) and F1 dropped from 84.1 % in the source domain to 66.3% in the target domain, the CORAL DA and adversarial DA improved it to 76.8% and 75.9% respectively. These improvements were over a half of the performance degradation, indicating the usability of our methods.