Abstract
This paper addresses the problem of optimal alignment of non-rigid surfaces from multi-view video observations to obtain a temporally consistent representation. Conventional non-rigid surface tracking performs frame-to-frame alignment which is subject to the accumulation of errors resulting in drift over time. Recently, non-sequential tracking approaches have been introduced which re-order the input data based on a dissimilarity measure. One or more input sequences are represented in a tree with reducing alignment path length. This limits drift and increases robustness to large non-rigid deformations. However, jumps may occur in the aligned mesh sequence where tree branches meet due to independent error accumulation. Optimisation of the tree for non-sequential tracking is proposed to minimise the errors in temporal consistency due to both the drift and jumps. A novel cluster tree enforces sequential tracking in local segments of the sequence while allowing global non-sequential traversal among these segments. This provides a mechanism to create a tree structure which reduces the number of jumps between branches and limits the length of branches. Comprehensive evaluation is performed on a variety of challenging non-rigid surfaces including faces, cloth and people. This demonstrates that the proposed cluster tree achieves better temporal consistency than the previous sequential and non-sequential tracking approaches. Quantitative ground-truth comparison on a synthetic facial performance shows reduced error with the cluster tree.