DARN: Crowd Counting Network Guided by Double Attention Refinement

Although great progress has been made in crowd counting, accurate estimation of crowd numbers in high-density areas and full mitigation of the interference of background noise remain challenging. To address these issues, we propose a method called Double Attention Refinement Guided Counting Network (DARN). DARN introduces an attention-guided feature aggregation module that dynamically fuses features extracted from the Transformer backbone. By adaptively fusing features at different scales, this module can estimate the crowd for high-density areas by restoring the lost fine-grained information. Additionally, we propose a segmentation attention-guided refinement method with multiple stages. In this refinement process, crowd background noise is filtered by introducing segmentation attention maps as masks, resulting in a significant refinement of the foreground features. The introduction of multiple stages can further refine the features by utilizing fine-grained and global information. Extensive experiments were conducted on four challenging crowd counting datasets: ShanghaiTech A, UCF-QNRF, JHU-CROWD++, and NWPU-Crowd. The experimental results validate the effectiveness of the proposed method.