Human-Centric Image Categorization Based on Poselets

In daily life, one kind of images are common, in which many people are present and performing certain activities. We call these images human-centric images. As the number of such images gets increasingly larger, to organize and access them efficiently becomes urgent. Since the categories of human-centric images are determined by human activities in the images, in this paper, we propose to classify human-centric images by analyzing poses of all humans in them. Specifically, first, we introduce the notion of poselets, which represent parts of poses of humans and a method to detect human based on the poselets. Given a human-centric image, to determine its category, we use the poselets and the human detection method to detect all possible poselet activations in it and create a statistical representation of the poses of humans in the image. Additionally, we also investigated the influence of contextual information on the categorization of human-centric images. Finally, for evaluating the human-centric image categorization method, five categories of human-centric images are collected from the internet and used for experiments. Experiment results show that the poselet distribution representations are more suitable for representing human-centric images than the popular bag of visual words method.