.png){kind=icon}
.png){kind=icon}
Return
STCC: Scale-aware transformer for crowd counting
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.114992.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In order to improve counting accuracy under different crowd density scenarios, a novel multi-scale aware crowd counting network named STCC is proposed in this paper. Specifically, it adopts Biformer-small with dynamic sparse attention for feature extraction, which automatically filters and retains key-value pairs for critical regions. Furthermore, an element-wise multi-scale token aware enhancement module is designed to refine feature maps of different scales, which enables our model to concentrate on features more relevant to crowd counting. Inspired by U-MixFormer, a decoder called U2-MixFormer is proposed. It not only enhances the model’s perception of spatial positions, but also further improves its ability to capture and integrate cross-scale features. Extensive experiments on six crowd counting datasets demonstrate that STCC has strong competitiveness, achieving state-of-the-art performance on the ShanghaiTech A and JHU-Crowd++ datasets.
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W