Hierarchical Shape Matching for Temporally Consistent 3D Video

Abstract

In this paper we present a novel approach for temporal alignment of reconstructed mesh sequences with non-rigid surfaces to obtain a consistent representation. We propose a hierarchical scheme for non-sequential matching of frames across the sequence using shape similarity. This gives a tree structure which represents the optimal path for alignment of each frame in the sequence to minimize the change in shape. Non-rigid alignment is performed by recursively traversing the tree to align all frames. Non-sequential alignment reduces problems of drift or tracking failure which occur in previous sequential frame-to-frame techniques. Comparative evaluation on challenging 3D video sequences demonstrates that the proposed approach produces a temporally coherent representation with reduced error in shape and correspondence.