A speed up efficient match kernel for visual estimation and detection

Low level local descriptors are so successful for their excellent performance in image correspondence matching problems. However, due to the intrinsic nature of high computational complexity, their areas of application shrink to a limited ground. In this paper, we address visual pedestrian detection and human pose estimation with a learning based feature representation from a speed up efficient match kernel, by mainly focusing on what image feature contributes to visual detection and human pose estimation, and to how much extent the learning process actually matters. When dealing with estimation accuracy combined with state-of-art prediction methods, we would like to testify that the modified patch level image representation could boost the accuracy for the final 3D pose prediction in monocular video images by information from image structural correlation to characterize image context, especially for human body shapes and motions. We have compared our proposed method with classic global features such as HOG, and conducted comprehensive experiments in detection and human pose estimation tasks on benchmark datasets. Final evaluation result verifies competitive discriminatory effectiveness and distinctiveness for our proposed feature in such visual tasks.