Counting Beyond Domains: Toward Alignment in Unsupervised Domain Adaptation in Remote Sensing Object Counting

The variability in altitudes, geographical background, and weather conditions across datasets degrades deep neural network (DNN) object-counting performance. Unsupervised domain adaptation (UDA) is an effective solution to bridge the domain gap but still suffers from insufficient and inaccurate alignment. To address this issue, we propose counting beyond domains (CBD), an alignment framework that comprehensively narrows domain gaps and effectively suppresses noise in pseudolabels for UDA. It is achieved by taking three key factors—scale, semantics, and style—into consideration. Specifically, we introduce: 1) a text-guided scale alignment (TGSA) module to perceive and align scales of objects between source and target domains with pretrained models; 2) a robust semantic alignment (RSA) module to select high-quality pseudolabels with a coarse-to-fine strategy; and 3) an adaptive instance normalization (AdaIN)-based style alignment (ASA) module to further narrow style variations across datasets using AdaIN. CBD provides a comprehensive and robust framework that overcomes misalignments due to scale and style variations, as well as pseudolabel noise. Comprehensive experiments validate the effectiveness of our approach, demonstrating its strong adaptability across various datasets.