Scalable Receptive Field GAN: An End-to-End Adversarial Learning Framework for Crowd Counting

摘要

Crowd counting is challenging for unrestricted open outdoor and diverse scenes. To address large variety of perspective, density distribution and clutter problems, a novel end-to-end deep generative adversarial framework with scalable receptive field (SRFGAN) is proposed for obtaining high quality density estimation in this paper. Specifically, our generator adopts an encoder-decoder network with residual blocks to achieve multi-scale features due to scalable receptive fields which adapts to different scale crowd distribution. We also explore a spatial global pooling layer to acquire image-level prior representation which helps to tackle severe perspective distortion and background clutter. Besides, feature matching loss and adversarial loss are combined via a joint training scheme, which helps to improve the quality of generated density map. Experiment results on ShanghaiTech and UCF_CC_50 datasets illustrate the superior effectiveness.