Non-rigid Shape Registration Using Similarity-Invariant Differential Coordinates

We present a method of modeling the three-dimensional shape of a deforming object from a time-series of range images. A non-rigid variant of ICP algorithm is proposed to solve the correspondence between the range images. The deformation is modeled by the combination of piecewise similarity transformation using differential coordinates, and solved stably and efficiently by alternating minimization of a quadratic objective function. The performance of the proposed method is demonstrated by conducting two modeling tasks. First, the complete shapes of human hand and full-body are reconstructed by integrating the range images acquired by a single depth camera. Second, the shape completion of four-dimensional scan data is carried out by fitting a template model to the point clouds by extending our registration technique in the spatiotemporal domain.