Human centric object detection in highly crowded scenes

In this paper, we propose to detect human centric objects, including face, head shoulder, upper body, left body, right body and whole body, which can provide essential information to locate humans in highly crowed scenes. In the literature, the approaches to detect multi-class objects are either taking each class independently to learn and apply its classifier successively or taking all classes as a whole to learn individual classifier based on sharing features and to detect by step-by-step dividing. Different from these works, we consider two issues, one is the similarities and discriminations of different classes and the other is the semantic relations among them. Our main idea is to predict class labels quickly using a Salient Patch Model (SPM) first, and then do detection accurately using detectors of predicted classes in which a Semantic Relation Model (SRM) is proposed to capture relations among classes for efficient inferences. SPM and SRM are designed for these two issues respectively. Experiments on challenging real-world datasets demonstrate that our proposed approach can achieve significant performance improvements.