IMD-Net: A Deep Learning-Based Icosahedral Mesh Denoising Network

In this work, we propose a novel denoising technique, the icosahedral mesh network (IMD-Net) for closed genus-0 meshes. IMD-Net is deep neural that produces denoised in single end-to-end pass, preserving and emphasizing natural object features process. A preprocessing step, exploiting homeomorphism between sphere, remeshes an irregular using regular structure of frequency subdivided icosahedron. Enabled by gauge equivariant convolutional layers arranged residual U-net, denoises remeshing invariant to global transformations as well local feature constellations orientations, doing so with computational complexity traditional conv2D kernel. The equipped carefully crafted loss function leverages differences positional, normal curvature fields target noisy numerically stable fashion. first, two large shape datasets commonly used related fields, ABC xmlns:xlink="http://www.w3.org/1999/xlink">ShapeNetCore , are introduced evaluate denoising. IMD-Net’s competitiveness existing state-of-the-art techniques established both metric evaluations visual inspection models. Our code publicly available at https://github.com/jjabo/IMD-Net .