Meshtryoshka: Differentiable Rendering of Real-World Scenes via Mesh Rasterization

Problem Statement
Differentiable rendering has become a cornerstone for 3D reconstruction and novel view synthesis, yet state-of-the-art methods predominantly rely on non-mesh representations (e.g., neural radiance fields, 3D Gaussian splatting) or specialized differentiable renderers. While 3D surface meshes are the industry standard for portability, hardware acceleration, and downstream graphics workflows, existing differentiable mesh rendering techniques are largely limited to object-centric scenes. They often require foreground-background masks, struggle with unbounded real-world environments, and typically depend on custom CUDA implementations or stochastic gradient estimation to handle the non-differentiability of standard rasterizers. Consequently, there is no existing method that simultaneously achieves photorealistic results on real-world scenes and directly optimizes a mesh representation using conventional graphics tools.

Methodology: Meshtryoshka
The authors introduce Meshtryoshka, a differentiable rendering framework that optimizes a 3D mesh representation of unbounded, real-world scenes using an off-the-shelf, non-differentiable triangle rasterizer. The core innovation lies in decoupling the optimization of vertex positions from the rasterization process.

1. Scene Parameterization and Mesh Extraction:
   The scene is represented by a sparse grid of signed distance function (SDF) values and spherical harmonics coefficients for view-dependent color. During each optimization step, the method extracts multiple level sets (iso-surfaces) from the SDF using a differentiable implementation of the Marching Cubes algorithm. This produces a "Matryoshka" of nested, non-intersecting triangle mesh shells. Each vertex in these shells is assigned a signed distance value and spherical harmonics coefficients via interpolation weights derived from the Marching Cubes process.

2. Non-Differentiable Rasterization & Deferred Rendering:
   Instead of implementing a custom differentiable ray-tracer, the method rasterizes each mesh shell independently using a standard, non-differentiable rasterizer. The rasterizer outputs per-pixel triangle IDs and barycentric coordinates. These are used to perform differentiable interpolation of the per-vertex SDF values and spherical harmonics coefficients in image space, effectively mimicking a fragment shader.
   To approximate volume rendering, the method assigns specific transmittance values ($T_i$) to each shell. The innermost shell is set to near-zero transmittance (opaque), ensuring it occludes all subsequent shells. The final image is constructed by alpha-compositing the rendered shells. Gradients flow from the rendering error back through the alpha-compositing process and the interpolation step to update the underlying SDF and color parameters, indirectly optimizing the mesh geometry.

3. Scalability to Unbounded Scenes:
   To handle real-world, unbounded scenes, the authors employ several strategies:

   • Sparsity: A sparse Marching Cubes implementation is used, storing parameters only for active voxels and helper voxels, drastically reducing memory usage compared to dense grids.
   • Coarse-to-Fine Optimization: The grid is initialized at low resolution and progressively subdivided, pruning empty space and refining geometry only where necessary.
   • Non-Cubic Grids: The scene is divided into a foreground (axis-aligned cube) and a background. The background is modeled using six truncated frustums with exponentially scaled vertices. This allows for efficient allocation of resolution, placing fewer parameters in distant regions while maintaining voxel aspect ratios suitable for Marching Cubes.

4. Regularization:
   To ensure stable optimization, the method employs a Laplacian smoothness loss on the SDF values (rather than an Eikonal loss, which can cause checkerboard artifacts in explicit grids) and an L1 penalty to encourage sparsity in unobserved regions. Additionally, the learning rate for non-DC spherical harmonic components is reduced to prevent the model from "cheating" geometry to fit specific views.

Key Contributions

• A Novel Framework: Meshtryoshka is the first differentiable mesh rendering method capable of reconstructing unbounded, real-world 3D scenes without requiring foreground masks.
• Compatibility with Standard Tools: The method demonstrates that high-quality differentiable rendering is possible using off-the-shelf, non-differentiable rasterizers, eliminating the need for custom differentiable renderers or stochastic gradient estimators.
• High-Performance Implementations: The authors contribute optimized implementations of classic graphics algorithms, including a sparse Marching Cubes algorithm and a software triangle rasterizer capable of rendering over 100 million triangles simultaneously.
• Direct Mesh Optimization: Unlike methods that extract meshes post-hoc from volumetric representations, Meshtryoshka directly optimizes the mesh representation, ensuring the optimization-stage quality matches the final mesh rendering.

Results
The method was evaluated on the NeRF-Synthetic (object-centric) and Mip-NeRF 360 (real-world) datasets.

• Object-Centric Scenes: On NeRF-Synthetic, Meshtryoshka performs competitively with state-of-the-art differentiable surface rendering methods like NeuS2 and outperforms prior mesh-based methods (nvdiffrec with DMTet/FlexiCubes) without requiring visibility masks.
• Real-World Scenes: On Mip-NeRF 360, the method successfully reconstructs challenging, unbounded scenes with high visual quality, capturing fine details such as bicycle spokes and tree leaves. While it lags slightly behind top-tier non-mesh methods like Zip-NeRF and 3D Gaussian Splatting in quantitative metrics (PSNR), it significantly outperforms previous mesh-based approaches, which failed to produce coherent reconstructions in these settings.
• Ablation Studies: Experiments confirm the necessity of the exponential vertex positioning, sparsity, regularizers, and the use of multiple shells for achieving high-quality results.

Significance and Claims
The paper posits that Meshtryoshka suggests a new path for differentiable rendering: it may be possible to achieve competitive 3D reconstruction and novel view synthesis by relying entirely on the conventional computer graphics rendering stack (standard rasterizers and mesh representations) rather than specialized differentiable renderers or non-mesh primitives. The authors claim that their work bridges the gap between the portability of meshes and the performance of modern differentiable rendering, enabling direct optimization of meshes for real-world scenes. However, the authors remain modest regarding limitations, acknowledging that their method currently produces more triangles than primitive-based approaches, has slower training times than fast NeRF/Gaussian methods, and can occasionally generate "floater" artifacts in sparsely observed regions.