Learning scalable Omni-scale distribution for crowd counting

Abstract

Crowd counting is challenged by large appearance variations of individuals in uncontrolled scenes. Many previous approaches elaborated on this problem by learning multi-scale features and concatenating them together for more impressive performance. However, such a naive fusion is intuitional and not optimal enough for a wide range of scale variations. In this paper, we propose a novel feature fusion scheme, called Scalable Omni-scale Distribution Fusion (SODF), which leverages the benefits of different scale distributions from multi-layer feature maps to approximate the real distribution of target scale. Inspired by Gaussian Mixture Model that surmounts multi-scale feature fusion from a probabilistic perspective, our SODF module adaptively integrate multi-layer feature maps without embedding any multi-scale structures. The SODF module is comprised of two major components: an interaction block that perceives the real distribution and an assignment block which assigns the weights to the multi-layer or multi-column feature maps. The newly proposed SODF module is scalable, light-weight, and plug-and-play, and can be flexibly embedded into other counting networks. In addition, we design a counting model (SODF-Net) with SODF module and multi-layer structure. Extensive experiments on four benchmark datasets manifest that the proposed SODF-Net performs favorably against the state-of-the-art counting models. Furthermore, the proposed SODF module can efficiently improve the prediction performance of canonical counting networks, e.g., MCNN, CSRNet, and CAN.