Optimal, Non-Rigid Alignment for Feature-Preserving Mesh Denoising

We present a simple, yet effective method for non-rigid alignment of point clouds. Our focus lies on developing a practical approach that allows us to do efficient, multi-way alignment of millions of points such as those produced by structured light scanners. Starting from an initial displacement field over the combined point cloud, our solution relies on an iterative smoothing scheme on the neighborhood graph of each sub-scan, reducing the Dirichlet energy of the displacement field. We compare a number of schemes for computing the initial displacement field, ranging from estimating the Laplacian of the combined point clouds to more traditional measures such as the point to point distance or point to plane distance.