One-shot adaptation for cross-domain semantic segmentation in remote sensing images

Abstract

Contemporary cross-domain remote sensing (RS) image segmentation has been successful in recent years. When the target domain data becomes scarce in some realistic scenarios, the performance of traditional domain adaptation (DA) methods significantly drops. In this paper, we tackle the problem of fast cross-domain adaptation by observing only one unlabeled target data. To deal with dynamic domain shift efficiently, this paper introduces a novel framework named Minimax One-shot AdapTation (MOAT) to perform cross-domain feature alignment in semantic segmentation. Specifically, MOAT alternately maximizes the cross-entropy to select the most informative source samples and minimizes the cross-entropy of obtained samples to make the model fit the target data. The selected source samples can effectively describe the target data distribution using the proposed uncertainty-based distribution estimation technique. We propose a memory-based feature enhancement strategy to learn domain-invariant decision boundaries to accomplish semantic alignment. Generally, we empirically demonstrate the effectiveness of the proposed MOAT. It achieves a new state-of-the-art performance on cross-domain RS image segmentation for conventional unsupervised domain adaptation and one-shot domain adaptation scenarios.