Global Illumination using Photon Ray Splatting

Robert Herzog, Vlastimil Havran, Shinichi Kinuwaki, Karol Myszkowski, Hans-Peter SeidelComputer Graphics Forum 2007 · DOI

TL;DR

This paper introduces "Photon Ray Splatting," a novel global illumination technique that improves rendering quality and efficiency by estimating light distribution directly in ray space, rather than on scene surfaces. This approach effectively reduces common rendering artifacts like boundary and topological bias, leading to more robust and visually pleasing results, especially for complex geometries.

What Problem Does This Paper Address?

Rendering realistic global illumination, which simulates how light bounces around a scene, is computationally expensive. Traditional methods like photon mapping often produce artifacts such as boundary bias (incorrect lighting near object edges) and topological bias (errors on complex surfaces). While these methods can be improved with techniques like final gathering, they significantly increase computation time. This paper aims to provide a faster, yet high-quality, global illumination solution that mitigates these artifacts without a prohibitive performance cost.

Methods

The core innovation is Photon Ray Splatting. Instead of gathering photon hit points on surfaces (as in standard photon mapping), this method estimates light density directly in "ray space." This means that a photon's contribution is considered along its entire path, not just at its intersection point.

Key techniques include:

  • Density Estimation in Ray Space: This is the central idea, decoupling light estimation from surface geometry and reducing bias.
  • KD-Tree for Eye Samples: An efficient data structure is used to quickly find relevant "eye samples" (points where light from the camera hits the scene) for photon splatting.
  • Bandwidth Selection: A novel heuristic is introduced to adapt the "splatting kernel width" (how far a photon's influence extends) based on the photon's path probability density. This helps preserve sharp details while reducing noise.
  • Spherical Harmonics (SH) Basis: For handling more complex lighting and BRDFs (how surfaces reflect light), the method can represent illumination and BRDFs using SH coefficients, allowing for efficient processing and filtering.
  • Radiance Caching: The algorithm is extended to incorporate radiance caching, a technique that speeds up rendering by storing and interpolating illumination samples. This is adapted to work with the SH basis and photon ray splatting.
  • Image Space Filtering: A 2D filter is applied in image space to reduce noise in the rendered images, preserving discontinuities.

Datasets

Not explicitly discussed in the paper, but the authors evaluate their method on several scenes: APPARTMENT, CORNELL BOX, CORNER ROOM, SIBENIK, CONFERENCE, and a custom "glossy APPARTMENT" scene. These scenes are used to demonstrate the algorithm's performance and visual quality under various lighting conditions and geometric complexities.

Key Insights and Ideas

  • Ray Space is Key: Estimating light in ray space, rather than just at surface intersection points, is a fundamental shift that directly addresses biases inherent in surface-based methods.
  • Adaptive Bandwidth Matters: Dynamically adjusting the influence of photons based on their path properties (like how many bounces they've had) is crucial for balancing detail preservation and noise reduction.
  • SH Basis for Efficiency: Using spherical harmonics provides a compact and efficient way to represent directional lighting, especially for moderately glossy surfaces.
  • Radiance Caching Synergy: Combining photon ray splatting with radiance caching offers a powerful way to achieve high-quality global illumination with significantly reduced rendering times.

Key Results

  • Reduced Bias: The method effectively eliminates boundary and topological bias, leading to more accurate lighting near object edges and on complex surfaces.
  • Improved Visual Quality: Compared to standard photon mapping, the results show better visual quality, particularly for indirect illumination and in scenes with complex geometry.
  • Efficiency Gains: The proposed method, especially when combined with radiance caching and filtering, achieves competitive rendering times, often outperforming traditional photon mapping for similar quality. For example, the APPARTMENT scene rendered with 500,000 photons took 15 seconds with their method (including caching and filtering).
  • Scalability: The algorithm scales well with image resolution and the number of photons.

Limitations and Future Work

  • Visibility Neglect: A primary limitation acknowledged is the neglect of visibility within the splatting footprint, which can lead to light leakage and inaccurate high-frequency shadows, especially for direct lighting.
  • Parameter Tuning: The method requires tuning several parameters for bandwidth selection (smoothness coefficient C, bandwidth sensitivity S, clamping parameter R) and for discontinuity buffer segmentation.
  • Glossiness Adaptability: The current radiance map implementation is not fully adaptive to the "glossiness" of surfaces.
  • Temporal Coherence: The authors suggest extending the method to the temporal domain by reusing radiance information from previous frames to further improve efficiency and reduce noise.
  • Visibility Approximation: Future work includes incorporating partial visibility information during ray traversal to improve accuracy for direct lighting and shadows.
  • Distributed Rendering: The algorithm is amenable to distributed setups, which could further accelerate rendering.

Summary updated

BibTeX
@article{herzog2007global,
  author = {Robert Herzog and Vlastimil Havran and Shinichi Kinuwaki and Karol Myszkowski and Hans-Peter Seidel},
  title = {{Global Illumination using Photon Ray Splatting}},
  journal = {Computer Graphics Forum},
  year = {2007},
  doi = {10.1111/j.1467-8659.2007.01073.x},
}
Figure