3D Human Pose Recovery from Monocular Images via Effi cient Visual Feature Selection

In this paper we propose a new examplar-based approach to recover 3D human poses from monocular images. Given the visual feature of each frame, pose retrieval is first conducted in th e examplar database to find relevant pose candidates. Then, d ynamic programming is applied on the weighted set of candidates to impose temporal coherence and recover a continuous pose sequence. We made two contributions within this framework. First, we propose to use an effi cient feature selection algorithm to select the optimal visual feature components. The feature selection task is formulated as a trace-ratio objective which measures the score of the selected feature component subset, and the objective is effi ciently optimized to get the global optimum. The selected components are used instead of the original visual feature to improve th e accuracy and effi ciency of pose recovery. As second contribution, we propose to use sparse representation to retrieve the pose ca ndidates, where the measured visual feature is expressed as a sparse linear combination of the labeled samples in the database. Sparse representation ensures that semantically similar po ses have larger probability to be retrieved. The effectiveness of our approach has been validated quantitatively through extensive evaluations on both synthetic data and real data, and qualitatively by insp ecting the results of the real time system we have implemented.