MSGSA: Multi-Scale Guided Self-Attention Network for Crowd Counting

Abstract

The use of convolutional neural networks (CNN) for crowd counting has made significant progress in recent years; however, effectively addressing the scale variation and complex backgrounds remain challenging tasks. To address these challenges, we propose a novel Multi-Scale Guided Self-Attention (MSGSA) network that utilizes self-attention mechanisms to capture multi-scale contextual information for crowd counting. The MSGSA network consists of three key modules: a Feature Pyramid Module (FPM), a Scale Self-Attention Module (SSAM), and a Scale-aware Feature Fusion (SFA). By integrating self-attention mechanisms at multiple scales, our proposed method captures both global and local contextual information, leading to an improvement in the accuracy of crowd counting. We conducted extensive experiments on multiple benchmark datasets, and the results demonstrate that our method outperforms most existing methods in terms of counting accuracy and the quality of the generated density map. Our proposed MSGSA network provides a promising direction for efficient and accurate crowd counting in complex backgrounds.