Perspective-guided Point Supervision Network for Crowd Counting

Crowd counting is critical for video surveillance and public safety. However, due to the impact of perspective effects, large-scale variations have become one of the main challenges affecting the counting performance. In this paper, we propose perspective-guided point supervision network (PPSNet), which embeds perspective information into a point-supervised network to better handle the scaling problem. First, we construct a point-supervised framework, which directly applies point annotations as learning targets to predict a set of candidate points and achieve accurate positioning of individuals. Then, PPSNet is constructed by stacking multiple perspective-guided fusion modules, which extract multi-scale features and fuse them guided by perspective attention based on perspective information. Perspective attention combines spatial attention and channel attention to learn long-range dependencies from different distance and capture the most important features among the scales. Experimental results on ShanghaiTech Part A and Part B, and WorldExpo'10 demonstrate that our proposed PPSNet outperforms the state-of-the-art methods.