Representing Appearance and Pre-filtering Subpixel Data in Sparse Voxel Octrees

Jacob Munkberg, Eric Heitz, Fabrice NeyretHigh Performance Graphics 2012 · DOI

TL;DR

This paper introduces a novel sparse voxel octree (SVO) representation and rendering technique that accurately captures and filters subpixel details for complex surfaces. It addresses the limitations of existing methods by accounting for correlations between surface attributes and visibility, enabling seamless, high-quality rendering across various scales and view-dependent effects.

What Problem Does This Paper Address?

Rendering highly detailed surfaces, especially those with intricate patterns and varying material properties, presents a significant challenge. Traditional graphics techniques often struggle with aliasing and visual artifacts when dealing with geometry at scales smaller than a pixel. Methods like MIPmapping, while useful for texture filtering, make simplifying assumptions about the correlation between surface attributes (like color or normal) and their visibility, leading to inaccuracies, particularly at grazing angles or with complex materials. Sparse Voxel Octrees (SVOs) offer a hierarchical way to represent geometry, but interpolating and filtering the appearance data within them to achieve smooth, artifact-free rendering, especially for view-dependent effects, remains an open problem. This paper tackles the issue of accurately representing and rendering subpixel surface appearance, including correlations between geometry, material properties, and visibility, to achieve high-quality, real-time graphics.

Methods

The core of this work lies in a new SVO representation and a differential cone tracing rendering approach.

  • Sparse Voxel Octree Representation: Each voxel in their SVO stores not just geometric information (like a signed distance field and its variance) but also statistical descriptors of the microscopic surface within it. This includes macroscopic normals, microgeometry roughness, and importantly, parameters for attributes like color. These attributes are represented with a mean value and a correlation parameter with depth, allowing for view-dependent effects.
  • Differential Cone Tracing: Instead of traditional ray tracing, the authors employ a differential cone tracing method. This involves breaking down the cone of light entering a pixel into smaller, successive cone elements. For each element, they pre-compute and then combine occlusion distributions and outgoing radiance. This approach allows for efficient integration of subpixel geometry and appearance.
  • Visibility and Attribute Correlation: A key innovation is the explicit modeling of correlations between surface attributes (like color) and their visibility. They leverage a Gaussian surface model and Smith's visibility formulation to account for how depth-dependent attributes change with viewing and lighting angles. This is crucial for realistic rendering of materials like fabrics or textured surfaces.
  • Pre-computation and Runtime: The system involves a pre-computation step to build the hierarchical SVO with filtered attributes. At runtime, the differential cone tracing efficiently queries this structure to render the scene, achieving near-constant computational complexity per pixel.

Datasets

The paper does not explicitly mention training datasets. For evaluation, they use:

  • Voxelized Marko Dabrovic's Sponza: A standard scene used for testing rendering techniques, voxelized to a high resolution (2048³ voxels).
  • Real-world materials: They demonstrate their method's ability to filter view-dependency on real materials, including anisotropic surfaces and textured objects, comparing their results against ground truth and simpler filtering methods.

Key Insights and Ideas

  • Correlation is King: The most significant insight is that accurately rendering detailed surfaces requires accounting for the correlation between surface attributes (color, roughness) and their visibility. Simple averaging or independent filtering of geometry and appearance leads to noticeable artifacts, especially at grazing angles.
  • Subpixel Details Matter: Even at scales smaller than a pixel, the distribution of surface details significantly impacts the final rendered image. Their SVO representation captures these subpixel statistics.
  • Differential Cone Tracing for Efficiency: Breaking down the rendering cone into differential elements allows for efficient pre-integration and scalable rendering, leading to near-constant per-pixel costs.
  • Seamless Zoom and Transitions: The hierarchical nature of the SVO and the pre-filtering approach enable smooth transitions when zooming in or out, avoiding the "popping" or aliasing artifacts common in other methods.
  • Material as a Function of Depth: Representing attributes as functions of surface depth, rather than just fixed values, is a powerful way to capture view-dependent material behavior.

Key Results

  • High-Quality Rendering: The method produces anti-aliased, coherent renderings with seamless transitions across scales and view-dependent effects, outperforming traditional MIPmapping and naive filtering approaches.
  • Real-time Performance: The algorithm achieves real-time performance (40-60 fps without shadows, 10-25 fps with shadows) on a GTX 560 GPU, with a cost per pixel that is nearly constant, even with complex silhouettes.
  • Accurate Material Filtering: They demonstrate accurate filtering of view-dependent color effects and emboss-to-shading transformations, closely matching ground truth for complex materials.
  • Efficient Memory Usage: The SVO representation is relatively compact, requiring approximately 15-20 bytes per voxel on average, which is comparable to or slightly more than existing SVO methods but offers significantly improved rendering quality.

Limitations and Future Work

  • Macroscopic B-rep Assumption: The current model assumes macroscopic surfaces are part of a B-rep object, meaning it doesn't explicitly handle thin objects or structures where internal parts are not visible.
  • Recursive Threads for Reflections: Implementing reflections and refractions would require recursive threads in CUDA, which is not currently supported in their implementation.
  • Complex Surfaces and Volumes: The model is primarily designed for surfaces. Extending it to handle more complex scenarios like volumes (e.g., grass, foliage, semi-transparent materials) with view-dependent effects is a significant area for future work.
  • Curvature of Coarse Surfaces: Accounting for the curvature of coarse surfaces remains a challenge.
  • Procedural Surface Enhancement: While their distance field representation allows for procedural surface enhancement, further exploration in this area is suggested.

Summary updated

BibTeX
@inproceedings{munkberg2012representing,
  author = {Jacob Munkberg and Eric Heitz and Fabrice Neyret},
  title = {{Representing Appearance and Pre-filtering Subpixel Data in Sparse Voxel Octrees}},
  booktitle = {High Performance Graphics},
  year = {2012},
  doi = {10.2312/eggh/hpg12/125-134},
}
Figure