Hierarchical block-based incomplete human mocap data recovery using adaptive nonnegative matrix factorization

Human motion capture (mocap) data has been widely utilized for realistic character animation, and the missing marker problem caused by occlusions or a marker falling off often results in an incomplete collection. In this paper, we present a hierarchical block-based incomplete human mocap data recovery approach by using adaptive nonnegative matrix factorization, which mainly consists of two layers: interior layer and exterior layer. In the interior layer, we first decompose the underling human skeleton model into five blocks and represent the whole human mocap data in terms of the block-based sub-chain motion clips, in which the moving trajectories of each sub-chain motion clip always share the approximately low-rank property. Then, an adaptive nonnegative matrix factorization method aiming at exploiting the low-rank structure and the nonnegativity constraint is presented to restore each incomplete sub-chain motion clip individually. In the exterior layer, we integrate the recovered sub-chain motion clips and further utilize the known entries within the raw mocap data to refine the corresponding restored data of same positions, whereby the whole incomplete human mocap data can be well recovered. Without any user assistance and the training priors, the experimental results have shown the reliable recovering performance in comparison with the state-of-the-art competing approaches. Graphical abstractDisplay Omitted HighlightsThe human mocap data is represented with five block-based sub-chain motion clips.The low-rank and nonnegativity properties of the mocap data are well exploited.An adaptive nonnegative matrix factorization method is proposed for motion recovery.Much better results have been achieved than state-of-the-art algorithms.