Ghost Attention Efficient Network for Crowd Counting

Abstract

Despite the great success of crowd counting networks on various applications, the redundant structure and excessive parameters still limit the deployment of counting models on mobile or embedded devices. In order to address this problem, we propose a lightweight density estimation architecture named Ghost Attention Efficient Network (GAEnet) for high-accuracy and real-time crowd counting. Firstly, the lightweight but powerful Ghost Extraction Network (GEN) are introduced to extract multi-level features. Meanwhile, we design the Cross-order Ghost Attention Module (CGAM) which is based on the Parameter-Sharing mechanism. CGAM aims at capturing higher-order discriminative information and obtaining wider receptive fields. In addition, the Weight-sharing Mask Density Producer (WMDP) are proposed to alleviate the influence of complex background on counting accuracy. Finally, we conduct extensive experiments on multiple crowd counting datasets. The experimental results have demonstrated the effectiveness of our GAEnet.