Point2Mesh: A Self-Prior for Deformable Meshes

摘要

In this paper, we introduce Point2Mesh, a technique for reconstructing asurface mesh from an input point cloud. Instead of explicitly specifying aprior that encodes the expected shape properties, the prior is defined automaticallyusing the input point cloud, which we refer to as a self-prior.The self-prior encapsulates reoccurring geometric repetitions from a singleshape within the weights of a deep neural network. We optimize thenetwork weights to deform an initial mesh to shrink-wrap a single inputpoint cloud. This explicitly considers the entire reconstructed shape, sinceshared local kernels are calculated to fit the overall object. The convolutionalkernels are optimized globally across the entire shape, which inherentlyencourages local-scale geometric self-similarity across the shape surface.We show that shrink-wrapping a point cloud with a self-prior converges to adesirable solution; compared to a prescribed smoothness prior, which oftenbecomes trapped in undesirable local minima. While the performance oftraditional reconstruction approaches degrades in non-ideal conditions thatare often present in real world scanning, i.e., unoriented normals, noise andmissing (low density) parts, Point2Mesh is robust to non-ideal conditions.We demonstrate the performance of Point2Mesh on a large variety of shapeswith varying complexity.