Re-evaluation of the CNN-based State-of-the-art Crowd-counting Methods with Enhancements

Abstract

Crowd counting has a range of applications and it is an important task that can help with the accident prevention such as crowd crushes and stampedes in political protests, concerts, sports, and other social events. Many crown counting approaches have been proposed in the recent years. In this paper we compare five deep-learning-based approaches to crowd counting, reevaluate them and present a novel CSRNet-based approach. We base our implementation on five convolutional neural network (CNN) architectures: CSRNet, Bayesian Crowd Counting, DM-Count, SFA-Net, and SGA-Net and present a novel approach by upgrading CSRNet with application of a Bayesian crowd counting loss function and pixel modeling. The models are trained and evaluated on three widely used crowd image datasets, ShanghaiTech part A, part B, and UCF-QNRF. The results show that models based on SFA-Net and DM-Count outperform state-of-the-art when trained and evaluated on the similar data, and the proposed extended model outperforms the base model with the same backbone when trained and evaluated on the significantly different data, suggesting improved robustness levels.