Crowd Counting in Images via DSMCNN

Abstract

Counting the crowd from a single image accurately is always a challenging task. In this work, we propose a Dilated Stacked Multi-column Convolutional Neural Network architecture for crowd density estimation in still single images. The model is composed of three columns of convolutional layers with sharing layers. We use smaller kernel and the dilated layer. We stack multifarious pooling layers and optimize the loss function. The DSMCNN model is an end-to-end and easy-trained system. Meanwhile, it shows robust for images with different perspective or crowd density. We demonstrate experiments on the ShanghaiTech dataset and the mall dataset.