Feature Reaggregation Network for Crowd Counting

In this paper, we propose a novel end-to-end network named Feature Reaggregation Network (FRNet) for crowd counting, which focuses on fusing the multi-scale features in the hierarchy for generating high-quality density maps. Two level and three level feature reaggregation modules are developed between the backbone network and the next feature extraction modules in the hierarchy so that the low-layer spatial feature and the high-layer semantic information can be multi-combined by element-wise addition. In addition, these two modules are placed behind the backbone network to extract global features. Furthermore, we also introduce the extra deformable mixture regression module which donates deformable convolution to extract features so that we can generate the high-quality estimated density map. We have evaluated our experiments on three popular crowd counting datasets (ShanghaiTech, UCF_CC_50 and UCF_QNRF datasets), and the experiments demonstrate that the superiority of the proposed method over the other excellent approaches.