Multilinear Motion Synthesis


We propose a practical method for interpolating motion samples with multilinear models. Our method first identifies a subspace of the given motion samples by analyzing the multimodal correlations of joints, time series, and control parameters, by which the redundancy of motion samples is fully eliminated. A new motion is then synthesized by interpolating the principal features of the samples using geostatistics. Our motion representation within a multilinear subspace successfully reduces the storage and computational costs of geostatistical motion interpolation while preserving the prediction accuracy. Adaptive selection of the number of principal features can control the level of detail of motion synthesis.


QuickTime movie
SCA 2006 video - part 1

QuickTime: 22.2 MB, with Audio
480×360, 2:23
QuickTime movie
SCA 2006 video - part 2

QuickTime: 20.0 MB, with Audio
480×360, 2:25


  1. Tomohiko Mukai and Shigeru Kuriyama, "Multilinear Motion Synthesis Using Geostatistics", ACM SIGGRAPH/Eurographics Symposium on Computer Animation 2006, Posters and Demos, pp.21-22, 2006.9.
    Paper (preprint, PDF: 40KB)
    Poster (PDF: 0.5MB)

Sample application

Parameterized punching motion

  • Windows application + Motion files (ZIP: 663KB)
        (DirectX End-User runtime October 2006 required)
  • References

    1. MATLAB Tensor Toolbox
    2. On the Best rank-1 and Rank-(R_1, R_2, ..., R_N) Approximation of Higher-Order Tensors
    3. A Multilinear Singular Value Decomposition
    4. Out-of-Core Tensor Approximation of Multi-Dimensional Matrices of Visual Data

    Return to TOP

    Last modified: 2006/11/23