Fine-Grained Crowd Counting

Current crowd counting algorithms are only concerned about the number of people in an image, which lacks low-level fine-grained information crowd. For many practical applications, total image is not as useful each sub-category. example, knowing waiting inline or browsing can help retail stores; standing/sitting restaurants/cafeterias; violent/non-violent police management. In this article, we propose counting, differentiates a into categories based on behavior attributes individuals (e.g. violent behavior) and then counts category. To enable research area, construct new dataset four real-world tasks: traveling direction sidewalk, standing sitting, line not, exhibiting not. Since appearance features different similar, challenge to effectively utilize contextual distinguish between categories. We two branch architecture, consisting density map estimation semantic segmentation branch. refinement strategies for improving predictions branches. First, encode information, feature propagation guided by prediction, eliminates effect background during propagation. Second, complementary attention model share Experiment results confirm effectiveness our method.