Perspective-Aware Loss Function for Crowd Density Estimation

Abstract

Estimation errors caused by perspective distortions are a long-standing problem in the domain of crowd counting. In this paper, we propose a novel loss function to allow filters in convolutional neural networks to learn features that are adaptive to the scale and perspective variation of individuals in crowd images. By exploring the crowd count error from regions close to the vanishing point of a perspective distorted image, we are able to penalize under-estimations. This is useful to train a network that is robust against perspective distortion for accurate density estimation. The proposed method is scene-independent and can be applied effectively to crowd scene with a variety of physical layout. Extensive comparative evaluations demonstrate that our proposed method achieves significant improvement over the state-of-the-art approaches on the challenging ShanghaiTech and UCF-QNRF datasets.