ManifoldPlus: A Robust and Scalable Watertight Manifold Surface Generation Method for Triangle Soups

TL;DR

<3-5 sentence high-level summary> ManifoldPlus tackles the challenging problem of turning artist-created triangle soups into watertight manifolds, addressing exterior-ambiguity and zero-volume structures that hinder existing methods. It combines a scalable octree-based occupancy surface extraction with a projection-based, inversion-free optimization (solved via Gauss-Seidel) that preserves sharp features and closely approximates a reference mesh. The approach yields robust, accurate, and scalable results on large-scale city models and standard benchmarks, outperforming state-of-the-art methods in correctness and efficiency. It also extends to scan-based reconstruction with improved geometric fidelity by avoiding reliance on input normals and maintaining manifold topology throughout the process.

Abstract

We present ManifoldPlus, a method for robust and scalable conversion of triangle soups to watertight manifolds. While many algorithms in computer graphics require the input mesh to be a watertight manifold, in practice many meshes designed by artists are often for visualization purposes, and thus have non-manifold structures such as incorrect connectivity, ambiguous face orientation, double surfaces, open boundaries, self-intersections, etc. Existing methods suffer from problems in the inputs with face orientation and zero-volume structures. Additionally most methods do not scale to meshes of high complexity. In this paper, we propose a method that extracts exterior faces between occupied voxels and empty voxels, and uses a projection-based optimization method to accurately recover a watertight manifold that resembles the reference mesh. Compared to previous methods, our methodology is simpler. It does not rely on face normals of the input triangle soups and can accurately recover zero-volume structures. Our algorithm is scalable, because it employs an adaptive Gauss-Seidel method for shape optimization, in which each step is an easy-to-solve convex problem. We test ManifoldPlus on ModelNet10 and AccuCity datasets to verify that our methods can generate watertight meshes ranging from object-level shapes to city-level models. Furthermore, through our experimental evaluations, we show that our method is more robust, efficient and accurate than the state-of-the-art. Our implementation is publicly available.

Figures (13)

• Figure 1: Illustration of manifold remeshing challenge. (a) Marching cubes cannot resolve sign conflicts at T-junctions. (b) There is natural ambiguity regarding exterior information. Non-manifold edges can be created at T-junctions by removing all exterior regions. (c) Our novel method addresses these problems.
• Figure 2: Mesh extraction example. The input is in (a) where problematic face orientation causes rendering artifacts in the zoomed-in planar region. (b) and (c) are manifold meshes extracted using our method at different resolutions. (d) is our final result based on (c) followed by an optimization step (Section \ref{['sec:approach-optim']}-\ref{['sec:approach-sharp']}), which is close to (a) but without less rendering artifacts.
• Figure 3: Non-manifold edge and vertex examples. (a) Non-manifold edges (red) appears when two diagonal voxels are occupied among the edge's four incident voxels. (b) Non-manifold vertex is shared by more than one group of connected triangles.
• Figure 4: Vertex $k$ are kept inside its surrounding polygon projected to local tangent planes specified by vertex normal of $i$ which equals or is adjacent to $k$.
• Figure 5: Illustration of optimal normal solution for Equation \ref{['eq:normal-update']}. Feasible solutions sits on the sphere surface inside the 3D polygon specified by linear constraints. Optimal solution is the normalization of the point $\hat{\mathbf{n}}_k$ inside the polygon that is closest to $\tilde{\mathbf{n}}_k$.

Theorems & Definitions (1)

• theorem 1