Watertight Mesh Optimization of Textured Objects

Nathaniel Burgdorfer, Philippos MordohaiarXiv 2026 · arXiv

TL;DR

MODL is a new framework that reconstructs high-quality 3D assets directly as watertight triangle meshes from simple RGB images. By combining Laplacian-preconditioned optimization with curvature-adaptive remeshing, it produces models that are both photorealistic and "production-ready"—meaning they are geometrically accurate, manifold, and suitable for physics or animation without any messy post-processing.

What Problem Does This Paper Address?

In the world of 3D reconstruction, there is a persistent "representation gap." Recent breakthroughs like 3D Gaussian Splatting (3DGS) produce stunningly photorealistic images but result in "triangle soups"—disconnected clouds of primitives that are useless for downstream applications like physics simulations or CAD.

Conversely, methods that focus on extracting clean geometry often struggle with fine texture details or require complex post-processing (like TSDF Fusion) that can introduce artifacts and blur appearance. Directly optimizing a connected triangle mesh is the "holy grail," but it is notoriously difficult: image-space gradients are noisy and spatially incoherent, often leading to "spiky" surfaces, self-intersections, and poor convergence.

Methods

MODL (Mesh Optimization via Differentiable Learning) bridges this gap by optimizing a watertight triangle mesh directly. The key technical pillars are:

  1. Laplacian-Preconditioned Optimization: Instead of moving vertices independently, MODL optimizes in differential coordinates. This propagates gradients across the mesh surface, ensuring that vertex updates are spatially coherent and smooth, effectively preventing the "spiky" artifacts common in direct mesh optimization.
  2. Curvature-Adaptive Remeshing: To capture fine details without wasting polygons on flat areas, the framework periodically remeshes the surface. It uses a curvature-dependent sizing field to split edges in high-detail regions (like the mane of a lion) and collapse them in simpler areas, all while maintaining well-shaped, near-equilateral triangles.
  3. Multi-view Photometric Constraints: Beyond standard image reconstruction losses (D-SSIM), MODL employs a multi-view Normalized Cross-Correlation (NCC) loss. This enforces geometric consistency by warping image patches from one view to another based on the exact mesh geometry, providing a powerful signal for surface accuracy.
  4. Skybox Gradient Sink: To avoid the need for manual segmentation masks, the authors place the object inside a "skybox" mesh. This skybox acts as a sink for background information, allowing the optimizer to focus exclusively on the foreground object.
The remeshing operations maintain triangle quality by splitting, collapsing, and flipping edges based on local curvature.

Datasets

The authors evaluated MODL on two primary real-world datasets:

  • DTU: A standard benchmark for multi-view stereo featuring objects with structured-light ground truth geometry.
  • Aria Digital Twin Catalog (DTC): A newer dataset providing high-resolution DSLR captures and detailed reference meshes with full 360-degree coverage.

Key Insights and Ideas

  • Unified Representation is Possible: You don't need to choose between a "rendering representation" and a "geometric representation." MODL shows that a single watertight mesh can handle both, provided the optimization strategy is robust.
  • Intrinsic Mesh Quality Matters: Most research focuses solely on Chamfer Distance (accuracy) or PSNR (rendering). MODL highlights the importance of intrinsic quality—metrics like the Self-Intersection Ratio (SIR) and Aspect Ratio (AR). A mesh that looks good but has thousands of self-intersections is practically useless for an engineer.
  • The Power of Preconditioning: The move from raw vertex optimization to Laplacian preconditioning is the "secret sauce" that makes direct mesh optimization stable enough for real-world, textured objects.
A radar chart comparing MODL to other methods across geometric, photometric, and intrinsic mesh quality metrics.

Key Results

MODL achieves a rare balance of performance across all metrics:

  • Geometric Accuracy: On the DTU dataset, MODL achieved an accuracy of 1.10mm, outperforming 2D Gaussian Splatting (1.15mm) and MeshSplatting (1.38mm).
  • Mesh Quality: MODL produced meshes with zero self-intersections (SIR) and zero non-manifold elements, whereas competing "splatting" methods often resulted in high rates of geometric defects.
  • Qualitative Fidelity: As seen in Figure 1, MODL successfully recovers deep concavities (like the eye sockets of a skull) and intricate textures that are often smoothed over by post-processing-heavy methods.
Figure 1. High-quality 3D asset reconstruction by MODL. Using only RGB images, camera poses, and sparse points as input, MODL jointly optimizes for geometric accuracy and appearance directly on a watertight triangle mesh, without requiring additional segmentation masks or estimated geometric priors. The models above, with geometry (left) and appearance (right), highlight MODL’s ability to recover renderable 3D models in a unified representation.

Limitations and Future Work

The primary limitation of MODL is topology. Because the framework deforms an initial mesh (the convex hull of sparse SfM points) and remeshing preserves the initial topology, it cannot change the "genus" of an object. For example, if you are reconstructing a donut but start with a sphere-like hull, MODL cannot currently "punch a hole" through the middle.

Future work aims to address this by incorporating topology-changing operations, allowing the framework to handle more complex, high-genus objects and entire scenes.

Summary updated

Abstract

When reconstructing high-quality 3D assets from images, existing approaches often excel at one objective while compromising another, and frequently overlook the intrinsic quality of the reconstructed surfaces. We present MODL, a framework that jointly optimizes geometry and view-dependent appearance directly on a watertight triangle mesh. Our approach combines Laplacian-preconditioned optimization for smooth vertex updates with curvature-adaptive remeshing to preserve fine details and well-shaped triangles. The triangle mesh is optimized subject to multi-view normalized cross-correlation constraints, but without segmentation, geometric supervision or alpha-blending. Evaluations on object reconstruction datasets demonstrate that MODL achieves competitive novel-view synthesis while producing geometrically accurate models with superior intrinsic mesh quality. Our results highlight the potential of direct mesh optimization to unify photorealistic rendering, geometric reconstruction, and practical 3D asset creation.
BibTeX
@misc{burgdorfer2026watertight,
  author = {Nathaniel Burgdorfer and Philippos Mordohai},
  title = {{Watertight Mesh Optimization of Textured Objects}},
  year = {2026},
  eprint = {2611.25414},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.25414},
}
Figure