Discriminative Human Pose Estimation Based on the Bandelet2 Image Descriptor

In this paper, we address the recovering from monocular images focusing on designing a novel image descriptor derived from the second generation Bandelet transformation, noted as Bandelet2, to tackle with estimation accuracy combined with state-of-art prediction methods. The proposed Bandelet2 image representation could boost the accuracy for the final 3D pose prediction in monocular video images by information from geometric flow to characterize image context, especially for human body shapes and motions. We have compared our image descriptor with classic ones as HOG, HMAX, laterally tested among different regression methods on standard Humaneva-I motion capture dataset and showed 3D reconstruction results. Final statistics verifies competitive discriminatory effectiveness and precision of Bandelet2 descriptor in estimating 3D human poses from monocular images.