Hirose Osamu
IEEE Trans Pattern Anal Mach Intell. 2021 Aug;43(8):2858-2865. doi: 10.1109/TPAMI.2020.3043769. Epub 2021 Jul 1.
Non-rigid point set registration is the process of transforming a shape represented as a point set into a shape matching another shape. In this paper, we propose an acceleration method for solving non-rigid point set registration problems. We accelerate non-rigid registration by dividing it into three steps: i) downsampling of point sets; ii) non-rigid registration of downsampled point sets; and iii) interpolation of shape deformation vectors corresponding to points removed during downsampling. To register downsampled point sets, we use a registration algorithm based on a prior distribution, called motion coherence prior. Using the same prior, we derive an interpolation method interpreted as Gaussian process regression. Through numerical experiments, we demonstrate that our algorithm registers point sets containing over ten million points. We also show that our algorithm reduces computing time more radically than a state-of-the-art acceleration algorithm.
非刚性点集配准是将表示为点集的形状转换为与另一个形状匹配的形状的过程。在本文中，我们提出了一种用于解决非刚性点集配准问题的加速方法。我们通过将非刚性配准分为三个步骤来加速它：i）点集的下采样；ii）下采样点集的非刚性配准；iii）对下采样过程中去除的点对应的形状变形向量进行插值。为了配准下采样点集，我们使用一种基于先验分布的配准算法，称为运动一致性先验。使用相同的先验，我们推导了一种解释为高斯过程回归的插值方法。通过数值实验，我们证明了我们的算法可以配准包含超过一千万个点的点集。我们还表明，我们的算法比一种先进的加速算法更能从根本上减少计算时间。
