Panet: Perspective-Aware Network with Dynamic Receptive Fields and Self-Distilling Supervision for Crowd Counting

Abstract

Crowd counting aims to learn the crowd density distribution and estimate the number of objects (e.g. persons) in images. The perspective effect, which significantly influences the distribution of data points, plays an important role in crowd counting. In this paper, we propose a novel perspective-aware approach called PANet to address the perspective problem in crowd counting. Based on the observation that the size of the objects varies greatly in an image due to the perspective effect, we propose a dynamic receptive fields (DRF) framework. The framework is able to adjust the receptive field by the dilated convolution parameters according to the perspective property of different local areas in the input image, which helps the model to extract more discriminative features for each local region. Furthermore, different from most previous works that use Gaussian kernels to generate the density map as the supervision information, we propose a self-distilling supervision (SDS) training method. The ground truth density maps are refined from the first training stage and the perspective information is distilled to the model in the second stage. With combination of the proposed DRF and SDS, our model can adapt to perspective effect and different density maps to achieve better accuracy. The extensive experimental results on ShanghaiTech Part_A and Part_B, UCF_QNRF, and UCF_CC_50 datasets demonstrate that our proposed PANet outperforms the state-of-the-art methods by a large margin.

Keywords: Crowd counting, Perspective effect, Dynamic receptive fields, Self-distillation