Stagewise Unsupervised Domain Adaptation With Adversarial Self-Training for Road Segmentation of Remote-Sensing Images

Road segmentation from remote sensing images is a challenging task with wide ranges of application potentials. Deep neural networks have advanced this field by leveraging the power large-scale labeled data, which, however, are extremely expensive and time-consuming to acquire. One solution use cheap available data train model deploy it directly process specific domain. Nevertheless, well-known domain shift (DS) issue prevents trained generalizing well on target In paper, we propose novel stagewise adaptation called RoadDA address DS in field. first stage, adapts features align source ones via generative adversarial (GAN) based inter-domain adaptation. Specifically, feature pyramid fusion module devised avoid information loss long thin roads learn discriminative robust features. Besides, intra-domain discrepancy domain, second an self-training method. We generate pseudo labels using generator divide easy split unlabeled hard road confidence scores. The adapted learning repeated progressively improve performance. Experiment results two benchmarks demonstrate that can efficiently reduce gap outperforms state-of-the-art methods.