Crowd counting via scale-adaptive convolutional neural network

TL;DR

This paper introduces SaCNN, a scale-adaptive CNN architecture for crowd counting that effectively handles scale variations and improves accuracy, especially in scenes with few pedestrians, outperforming existing methods.

Contribution

The paper proposes a novel scale-adaptive CNN with a fixed small receptive field backbone and a relative count loss, enhancing crowd counting accuracy across diverse scenes.

Findings

SaCNN outperforms state-of-the-art methods on multiple datasets.

The relative count loss improves performance in scenes with few pedestrians.

Extensive experiments validate the effectiveness of the proposed approach.

Abstract

The task of crowd counting is to automatically estimate the pedestrian number in crowd images. To cope with the scale and perspective changes that commonly exist in crowd images, state-of-the-art approaches employ multi-column CNN architectures to regress density maps of crowd images. Multiple columns have different receptive fields corresponding to pedestrians (heads) of different scales. We instead propose a scale-adaptive CNN (SaCNN) architecture with a backbone of fixed small receptive fields. We extract feature maps from multiple layers and adapt them to have the same output size; we combine them to produce the final density map. The number of people is computed by integrating the density map. We also introduce a relative count loss along with the density map loss to improve the network generalization on crowd scenes with few pedestrians, where most representative approaches perform poorly on. We conduct extensive experiments on the ShanghaiTech, UCF_CC_50 and WorldExpo datasets as well as a new dataset SmartCity that we collect for crowd scenes with few people. The results demonstrate significant improvements of SaCNN over the state-of-the-art.