Abstract

We propose a real-time method for simultaneously refining the reconstructed volume of a human body with loose-fitting clothing and identifying body-parts in it. Time-series volumes, which are acquired by a slow but sophisticated 3D reconstruction algorithm, with body-part la- bels are obtained o?ine. The time-series sample volumes are represented by tra jectories in the eigenspaces using PCA. An input visual hull recon- structed online is pro jected into the eigenspace and compared with the tra jectories in order to find similar high-precision samples with body- part labels. The hierarchical search taking into account 3D reconstruc- tion errors can achieve robust and fast matching. Experimental results demonstrate that our method can refine the input visual hull including loose-fitting clothing and identify its body-parts in real time.