Scale-equivariant Steerable Networks for Crowd Counting

Abstract

In the crowd counting based on density map estimation, the scale variance is the main factor causing the difficulty of crowd density map estimation. To handle this issue, we utilize the scale-equivariant steerable convolution and attention mechanism to build the network for the crowd counting. Meanwhile, we propose a method of generating high-precision pseudo bounding boxes about heads of crowd to generate the density maps based on the length and width of the pseudo bounding boxes which further obtain the scale information of the crowd. The density map improves the accuracy of density map estimation without changing our network configuration. In order to verify the effectiveness of scale-equivariant steerable convolution in crowd density map estimation, experiments are conducted on ShanghaiTech dataset and compared with other crowd counting networks. Experiments show that the proposed network structure and the generated density map are effective for the task of crowd counting based on density map estimation.