To Splat Straight with Crooked Points: Rendering Noisy Meshes and Point Clouds using Coherent Tangent Vector Fields

Abstract

Surface aligned splatting is a popular rendering technique to visualize reconstructed meshes and point clouds scanned from the real world. Such data typically presents some degree of noise that jeopardizes any attempt to render a perfectly smooth normal field and, more importantly, the estimated tangent vector fields are not locally continuous, thus affecting the overall visual quality. In this work, we compare two splat orientation techniques for rendering 3D noisy data, namely, the Covariance Matrix and the Householder formula. We evaluate both techniques using four publicly available meshes with synthetic noise, and four scanned point clouds with natural noise. Results indicate that the Householder technique is better suited for surface aligned splatting as it generates more coherent tangent vector fields, while Covariance Matrix reacts poorly to noise.