SinCount:Fourier Transform-Based Single Domain Generalization for Crowd Counting

Crowd counting plays a vital role in public safety and urban management. However, existing models often fail to generalize to unseen scenarios due to domain shifts. Domain Generalization (DG) can alleviate this issue, but most studies focus on classifification, while crowd counting under the Single-source DG (SDG) setting remains largely unexplored. In this paper, we propose SinCount, a novel SDG framework for crowd counting that integrates frequency-aware attention into an existing dual-branch architecture. Specififically, we extract high-frequency features from the density feature maps to generate spatial attention, which is applied back to the density branch to enhance fifine-grained details and domain-invariant representations. Meanwhile, low-frequency features are extracted from the classifification feature maps to produce channel attention, guiding the classifification branch to focus on semantic consistency and class-aware discrimination. We evaluate our method on multiple benchmark datasets and demonstrate that it achieves competitive results compared to state-of-the-art DG approaches.