Object-aware Conditional Alignment for Cross-domain Counting

Zhuonan Liang, Dongnan Liu, Jianan Fan, Yu Yao, Qiang Qu, Peng Fu, Weidong CaiOpenReview 2025 · Submission to ICLR 2025 · Rejected · OpenReview

Abstract

Object counting is an important task in computer vision with many real-world applications. In practical settings, factors such as lighting conditions and object density can vary dramatically, leading to distribution shifts then causing inaccurate counting. We found that existing domain adaptation (DA) methods cannot be directly applied to the counting task, as they usually assume changes across different domains are task-irrelevant and focus on utilizing domain-invariant features for prediction. However, in object counting tasks, changes in object density which could happen across domains are task-relevant and cannot be ignored. Therefore, applying existing DA methods to the counting task can ignore the information about density changes, resulting in unreliable counting. To address this limitation, we propose the Binary Alignment Network (BiAN). Unlike traditional DA methods that align distributions of entire image representations, BiAN segments objects of interest and aligns the distributions of the object-specific features across domains. This targeted alignment allows us to disregard irrelevant features, such as lighting conditions, while preserving essential information about changes in object density. We theoretically demonstrate that BiAN achieves superior adaptability in counting tasks by introducing conditional alignment—aligning features conditioned on the presence of objects. Extensive experiments on two distinct counting tasks and eight dataset combinations show that BiAN outperforms state-of-the-art methods.
BibTeX
@misc{liang2025object,
  author = {Zhuonan Liang and Dongnan Liu and Jianan Fan and Yu Yao and Qiang Qu and Peng Fu and Weidong Cai},
  title = {{Object-aware Conditional Alignment for Cross-domain Counting}},
  year = {2025},
  url = {https://openreview.net/forum?id=USGY5t7fwG},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2025},
}
Figure