Sparse geometric features for visual detection and estimation

Image feature is a crucial perquisite for computer vision community, such as visual pedestrian detection and human pose estimation. In this paper, we propose a novel image feature utilizing sparse geometric representation, to deal with image patterns that closely engaged in visual human detection and pose estimation problems. Also, the speed up optimization here ensures a promising performance for effective feature extraction. Combined with state-of-art learning methods, this paper indicates how the proposed image features could boost the representativeness for specific vision tasks from monocular video images by leveraging geometric flow information to characterize image context, especially for human body shapes and motions. We have compared our proposed method with classic global features such as HOG, HMAX, and conducted comprehensive experiments in detection as well as human pose estimation tasks on benchmark datasets. Final evaluation result verifies competitive discriminatory effectiveness and distinctiveness for our proposed feature in such vision tasks.