Bilateral counting network for single-image object counting

Abstract

This paper proposes a novel bilateral counting network to estimate the accurate and robust counting result for single-image object counting task. The proposed network is composed of two main components: the concentrated dilated pyramid module and dual-context extraction path. The concentrated dilated pyramid module extracts the multi-scale feature from the image to address the scale variant issue in object counting task via a pyramid structure and also uses a shortcut concentration to facilitate the back-propagation of the gradient so as to improve the counting performance. And the dual-context extraction path obtains different-level context related to the object counting task through convoluting and down-sampling the image different times. The concentrated dilated pyramid module and the dual-context extraction path are integrated to boost the final counting result. Extensive experiments on vehicle counting and crowd counting datasets including TRANCOS, Mall, Shanghaitech_A and WorldExpo’10 demonstrate the feasibility and effectiveness for the object counting task.