RHNet: Lightweight Dilated Convolutional Networks for Dense Objects Counting
Rui Yu, Xiangyang Xu, Yexiong Shen · 2019
In this paper, we propose a lightweight network called RHNet to achieve a faster counting method for dense objects counting. RHNet is composed of two major components. The first part contains four convolution layers and two max-pooing layers, which is designed to extract features mainly. The following second part is a structure named dilated special pyramid pooling, which is aimed at understanding the multiscale information. Compared to other published excellent networks, RHNet is a lightweight network because it has only 0.03 million parameters. Importantly, we demonstrate RHNet a decent accuracy on famous ShanghaiTech crowd dataset, WorldExpo'10 crowd dataset, UCF-QNRF dataset and an extended dataset with more kinds of dense objects.