An object counting network based on hierarchical context and feature fusion

Abstract

Object counting is a challenging task in computer vision. In this paper, we propose an object counting network based on hierarchical context and feature fusion called HFNet. HFNet comprises a hierarchical context extraction module and an end-to-end convolution neural network. The hierarchical context extraction module extracts hierarchical features to the main network as context cues, aiming to provide more information to improve counting performance. The main network adds the relatively lower but naturally high-resolution feature maps into higher but semantic feature maps, whose benefits are: one is to reduce the risk of losing detailed information during multi-convolutions; the other is to against the scale variations in this task due to the fusion operation of the multi-scale feature maps. Experiments demonstrate HFNet achieves competitive results on crowd counting including UCF_CC_50 dataset and ShanghaiTech dataset and on vehicle counting including TRANCOS dataset. The contrast experiments also verify the structure rationality of HFNet.