Source-linked AI summary
Elastic Triangle Splatting
Tian Shi, Shenhan Qian, Daniel Cremers
TL;DR
Existing neural rendering representations achieve high visual fidelity but remain poorly aligned with mesh-based graphics pipelines, and triangle-splatting kernel behavior has not been isolated systematically. The paper compares kernels across controlled mesh reconstruction and triangle-soup novel-view synthesis, then introduces an elastic kernel with bilateral gradient support and adaptive boundary opacity. The elastic kernel improves over existing kernels and achieves state-of-the-art performance in both tasks, while the shape-reconstruction pipeline remains limited by fixed topology.
Problem
Triangle-splatting methods use inconsistent kernels, while kernel optimization behavior has been insufficiently isolated beyond unstructured triangle-soup novel-view synthesis.
Method
The paper unifies kernel analysis across connected-mesh shape reconstruction and triangle-soup novel-view synthesis, introducing an elastic kernel with bilateral support and adaptive boundary opacity.
Results
The elastic kernel improves over existing kernels and achieves state-of-the-art performance in both shape reconstruction and novel-view synthesis.
Takeaways & Limitations
Kernel design, especially bilateral gradient support and adaptive boundary behavior, is important to triangle splatting’s optimization effectiveness across both tasks.
Takeaways & Limitations
The shape-reconstruction pipeline assumes fixed mesh topology, limiting it to object-centric settings.
Abstract
from arXiv · showhide
While neural rendering methods such as 3D Gaussian Splatting achieve remarkable visual fidelity, traditional polygonal meshes remain the backbone of established graphics pipelines. Triangle splatting bridges this gap by optimizing triangle primitives as differentiable splats, producing representations that are closer to mesh-based workflows. Central to these methods is the kernel function that softens triangle boundaries to propagate gradients to vertex positions. Existing triangle splatting methods make inconsistent choices of kernel functions, and analysis of these kernels' optimization behavior has been limited to unstructured triangle soups for novel-view synthesis. In this work, we consider triangle splatting as a generic tool for photometric optimization, comparing kernel properties through two complementary tasks: mesh optimization for shape reconstruction and triangle soup optimization for novel-view synthesis. Along with the analysis, we introduce an elastic kernel function that features bilateral gradient support across the boundary and an adaptive boundary value, which are shown to be essential for robust optimization. Under isolated comparison, our elastic kernel outperforms existing kernels on shape reconstruction and in the majority of novel-view synthesis benchmarks, demonstrating the importance of kernel design in the effectiveness and versatility of triangle splatting.
1. Introduction
Triangle splatting adapts triangle primitives for differentiable optimization while retaining compatibility with mesh-based graphics pipelines. This paper compares kernel behavior across mesh reconstruction and triangle-soup novel-view synthesis, introducing an elastic kernel that improves both tasks.
- Motivation: Triangle splatting uses differentiable triangle primitives to combine neural rendering quality with mesh-compatible representations.Optimized triangles can ultimately be rendered through standard rasterization without conversion from volumetric primitives.
- Motivation: Existing kernels differ in boundary support and are difficult to compare because prior evaluations use different pipelines and tasks.Triangle Splatting is single-sided, whereas 2DTS spreads color bilaterally across the boundary.
- Evaluation: Mesh optimization provides a controlled evaluation setting that isolates kernel effects from favorable initialization and adaptive densification in novel-view synthesis.The study uses a fixed-topology spherical mesh and compares against differentiable rasterization baselines such as Nvdiffrast.
- Evaluation: The paper evaluates kernel support, smoothness, decay, and boundary value across connected-mesh shape reconstruction and triangle-soup novel-view synthesis.The two tasks respectively optimize a connected mesh for geometric accuracy and independent triangles for photometric fidelity.
- Contribution: The elastic kernel combines wide bilateral gradient support with an adaptive boundary value to improve optimization across both settings.The paper reports improvements over Triangle Splatting and 2DTS, with state-of-the-art results in shape reconstruction and novel-view synthesis.
2. Related Work
Related work develops differentiable rendering for mesh optimization and neural scene reconstruction through implicit fields, Gaussian primitives, and triangle-based representations. These approaches progressively connect differentiable image formation with geometric surfaces and graphics workflows.
- Differentiable Rendering: Early differentiable-rendering methods approximate rasterization or silhouette gradients to enable gradient-based mesh optimization.OpenDR introduced approximate pipeline differentiation, while Neural Mesh Renderer used pixel-wise blurring for silhouette gradients.
- Neural Representations: NeRF, NeuS, and VolSDF use volumetric or implicit representations to reconstruct scenes while addressing surface ambiguity, smoothness, or topology.NeuS and VolSDF integrate volumetric rendering with signed distance functions.
- Gaussian Representations: 3DGS uses differentiable anisotropic Gaussian primitives for efficient real-time rendering, while later variants pursue principled densification or improved geometric reconstruction.2DGS constrains primitives to planar Gaussians, and SuGaR regularizes Gaussians against an underlying mesh surface.
3. Method
The method unifies triangle splatting across mesh reconstruction and triangle-soup novel-view synthesis, focusing on how kernel design controls gradient flow and boundary behavior. It introduces an elastic kernel with bilateral support and an adaptive boundary value, alongside regularized mesh optimization.
- Unified framework: The framework applies triangle splatting to both connected mesh optimization for shape reconstruction and triangle-soup optimization for novel-view synthesis.Both tasks use triangles as the rendering primitive, but differ in initialization, topology, and regularization.
- Kernel analysis: Kernel design determines how positional gradients are distributed around triangle boundaries, making support, smoothness, decay, and boundary value central comparison axes.The study compares single-sided and bilateral support together with fixed or adaptive boundary values under a unified distance parameterization.
- Kernel analysis: Single-sided kernels provide no exterior gradient, so triangles whose boundaries retreat beyond target regions can lose gradient flow and collapse.Their weight is zero outside the triangle, causing optimization to stall once the boundary moves past a target surface region.
- Kernel analysis: Bilateral kernels preserve interior and exterior gradient flow but fixed boundary weights can create edge artifacts and visual expansion when softened.For connected triangles, the fixed boundary value can leave transparent seams as the kernel is sharpened and bias geometry as softness increases.
- Elastic kernel: The elastic kernel combines bilateral gradient support with a boundary value that adapts with softness, supporting broad early movement and sharper final boundaries.As σ decreases, G(1) approaches one, reducing edge artifacts while retaining wide support when σ is large.
- Mesh optimization: Mesh optimization deforms a fixed-topology sphere using vertex positions, face appearance, opacity, and shared softness, while geometric regularizers preserve mesh validity.The objective combines L1 photometric loss with edge, normal-consistency, Laplacian, and softness regularization; the softness penalty prevents σ from collapsing to zero.
4. Experiments
Experiments evaluate the elastic kernel in mesh optimization and novel-view synthesis under unified comparisons. It improves geometric reconstruction, convergence, perceptual quality, and boundary sharpness relative to splatting-based alternatives.
- Evaluation Setup: The evaluation covers shape reconstruction on FAMOUS and novel-view synthesis on Mip-NeRF 360, Tanks&Temples, and Deep Blending.Shape reconstruction uses Chamfer Distance, while rendering quality uses PSNR, SSIM, and LPIPS.
- Shape Reconstruction: Our elastic kernel consistently outperforms splatting-based baselines on FAMOUS shape reconstruction under both λσ settings.The comparison uses shared mesh initialization, training schedule, and regularization, isolating optimization differences attributable to kernel design.
- Shape Reconstruction: The elastic kernel recovers sharper surfaces and finer geometric details than Triangle Splatting and 2DTS on FAMOUS.This qualitative comparison is consistent with the reported geometric and convergence improvements.
- Shape Reconstruction: The elastic kernel converges faster and achieves lower Chamfer Distance than Triangle Splatting and 2DTS throughout mesh optimization.Its reconstructed surfaces are cleaner and contain more accurate geometric details, while 2DTS still exhibits visible folds and cracks.
- Novel-View Synthesis: In novel-view synthesis, the method leads triangle-based baselines in SSIM and LPIPS across every benchmark, while the PSNR gap on Tanks&Temples is marginal.Performance differences are smaller than in shape reconstruction, with densification and pruning partly masking kernel behavior.
- Novel-View Synthesis: The optimized triangle soup renders with sharp, opaque boundaries, and sharper boundaries do not require increasing primitive count solely to compensate for boundary blur.The qualitative NVS results also show better visual details and fewer floaters and artifacts.
5. Conclusion
Elastic Triangle Splatting provides a unified framework for connected meshes and triangle soups, with an elastic kernel designed for robust optimization and clean boundaries. Experiments report consistent improvements and state-of-the-art performance across shape reconstruction and novel-view synthesis, while shape reconstruction remains limited by fixed topology and primitive density.
- Conclusion: Elastic Triangle Splatting unifies differentiable rendering for connected meshes and unstructured triangle soups.The elastic kernel combines bilateral gradient support with adaptive boundary opacity, producing clean boundaries as σ decreases.
- Conclusion: Experiments show consistent improvements over existing kernels and state-of-the-art performance in shape reconstruction and novel-view synthesis.
- Limitations: Fixed mesh topology limits shape reconstruction to object-centric settings.Complex or open scenes may require topology-variant optimization or hybrid representations.
- Limitations: Very fine geometric details may require many triangle primitives, creating a primitive-density trade-off similar to Gaussian splatting.
6. Technical Details
The technical details unify kernel formulations and preserve task-specific evaluation choices. Shape reconstruction uses controlled mesh refinement, while novel-view synthesis uses densification and pruning inherited from established triangle-splatting workflows.
- Kernel Formulation: All kernel functions are reformulated using barycentric distance x as a unified input, while their original forms are summarized separately.
- Kernel Formulation: The normalized variable s maps interior, boundary, and exterior points through s = 1 −x.Positive s denotes interior points, s = 0 the boundary, and negative s the exterior.
- Kernel Formulation: The 2DTS kernel is reparameterized so smaller σ yields sharper kernels, matching the comparison convention used for shape reconstruction.The original parameterization has the opposite relationship between σ and sharpness.
- Shape Reconstruction: Shape reconstruction adaptively subdivides large triangles by cosubdividing their longest shared edges, preserving topological consistency.Each triangle is processed at most once per subdivision step.
- Novel-View Synthesis: Novel-view synthesis clones, splits, and prunes triangle primitives during training, sampling sharper or more opaque triangles more often.The strategy concentrates capacity in high-error or under-covered regions.
- Experimental Setup: Shape reconstruction uses 21 FAMOUS objects with 32 training views and 8 test views per object, while NVS initializes triangles from SfM point clouds.Shape reconstruction starts from a normalized level-4 icosphere and runs for 30,000 iterations; NVS follows the original Triangle Splatting hyperparameters.
7. Additional Experiments
Additional experiments examine triangle filtering, regularization sensitivity, efficiency, and per-scene benchmark results. They indicate that small-triangle culling is unnecessary for NVS, extreme regularization can over-smooth meshes, and the elastic kernel adds modest training overhead while retaining comparable rendering throughput.
- Small Triangle Filtering: Not culling small triangles has negligible effect on novel-view synthesis performance.Small-triangle filtering can nevertheless harm mesh optimization by introducing rendering holes as subdivision makes the mesh finer.
- Regularization Sensitivity: The original regularization setting performs best, while large edge or Laplacian weights over-smooth the mesh and degrade Chamfer Distance.Moderate changes to normal and Laplacian terms preserve the same overall behavior.
- Runtime Efficiency: The elastic kernel introduces a modest training-time overhead while keeping triangle count and rendering throughput comparable to prior triangle kernels.The efficiency comparison reports optimization time, triangle count, FPS, and rendering time per 1 million triangles on Mip-NeRF 360.
- Per-Scene Results: Per-scene results are reported for Mip-NeRF 360, Tanks&Temples, and Deep Blending, including CD, PSNR, SSIM, and LPIPS metrics.