Crowd Counting via Single-Domain Generalization

To address the issue of “domain shift”, researchers have proposed numerous single-domain generalization methods for crowd counting. These methods primarily focus on feature alignment to directly extract domain-invariant features, thereby enhancing the model's generalization ability. However, we argue that in single-source domain scenarios, where the style differences in video images are minimal, the domain-invariant features extracted directly from the entire image often contain a significant amount of redundant background information. Based on this observation, we propose a crowd counting network via single domain generalization (Crowd-SDG). Specifically, to increase the diversity of the source domain, we design a block-style attention strategy. This strategy focuses on different regions of the same image and combines data augmentation techniques such as lighting enhancement, cropping, and stretching to obtain more diverse features. To extract domain-invariant features that contain only crowd-related information, we introduce a feature recognition module. This module first uses a contrastive language image pre-training to obtain image patches that contain only crowd features, then extracts the domain-invariant features related to the crowd via similarity computation. Additionally, we design a dynamic feature reconstruction module to effectively integrate domain-invariant features with regular features. This not only enhances the representation of domain-invariant information but also reduces the loss of fine-grained details. Extensive experimental results demonstrate that our method exhibits superior generalization performance across multiple datasets.