Dynamic Domain Distribution Alignment Through Online Optimal Transport

Unsupervised domain adaptation facilitates the transfer of knowledge from a labeled source domain to an unlabeled target domain, and its application in crowd counting is on the rise. The existing methods often lack in-depth exploration of inter-domain relationships, resulting in suboptimal alignment performance. To mitigate these challenges, we propose a Dynamic Domain Distribution Alignment Through Online Optimal Transport (OOTD) to align the domain-agnostic factors across domains dynamically. Our method unfolds in three phases: first, the fine-grained feature distributions of both the source and target domains are collected; second, the optimal transport algorithm is employed to establish a one-to-one mapping relationship between the source and target domains; and finally, our proposed patience-based early-exit mechanism is employed to iteratively and dynamically align the source and target domains. Our comprehensive experiments across six cross-domain benchmarks demonstrate that our method outperforms existing state of the art methods in adaptive crowd counting. Ablation studies and visualization results further validate the effectiveness of our proposed method.