Approximating dynamic global illumination in image space

Tobias Ritschel, Thorsten Grosch, Hans-Peter SeidelSI3D 2009 · DOI

TL;DR

This paper introduces Screen-Space Directional Occlusion (SSDO) and a single-bounce indirect illumination approximation that runs in image space. This allows for dynamic, high-resolution scenes to achieve more realistic lighting effects like colored shadows and indirect light bounces at real-time frame rates, significantly improving upon traditional ambient occlusion.

What Problem Does This Paper Address?

Achieving physically plausible global illumination (GI) in real-time for complex and dynamic scenes remains a significant challenge in computer graphics. Traditional methods often rely on approximations to meet performance demands. Ambient Occlusion (AO) is a popular approximation that darkens areas based on geometric occlusion, but it ignores the direction of incoming light and thus lacks directional shadows and color bleeding. Screen-space AO (SSAO) methods offer performance benefits by operating in image space but inherit AO's limitations. This paper aims to bridge this gap by extending SSAO to incorporate directional information and a single bounce of indirect light, all while maintaining real-time performance for dynamic scenes.

Methods

The core contribution is the generalization of SSAO to Screen-Space Directional Occlusion (SSDO) and the addition of one-bounce indirect illumination.

  • SSDO: Instead of just calculating a general occlusion factor, SSDO samples directions from a hemisphere around a pixel's normal. For each sample direction, it performs a screen-space visibility test by taking steps in that direction and checking if the sampled point is occluded by nearby geometry. This allows the method to account for the direction of incoming light, producing directional shadows and colored indirect lighting.
  • One-Bounce Indirect Illumination: This is achieved by using the direct lighting computed in the previous pass. For each sample point identified as an occluder in the SSDO step, its computed direct light is used as the "sender radiance" for a small patch on the surface. This radiance is then added to the target pixel, simulating a single bounce of indirect light.
  • Screen-Space Operation: Crucially, both SSDO and indirect illumination are computed entirely in image space. This means the computational cost is largely independent of the geometric complexity of the scene, making it suitable for dynamic and high-polygon count scenes.
  • Extensions for Quality: To address limitations of single-depth screen-space methods (e.g., missing occluders behind visible surfaces), the paper proposes two extensions:
  • Depth Peeling: Rendering multiple depth layers allows for a more accurate visibility test by considering geometry beyond the first visible surface.
  • Additional Cameras: Using multiple viewpoints can capture occluded regions and improve the detection of offscreen blockers.

Datasets

Not discussed in the paper. The paper uses example scenes for demonstration and evaluation, such as a factory scene and a Buddha model, but does not refer to standardized datasets.

Key Insights and Ideas

  • Screen-space is more than just occlusion: The paper demonstrates that the screen-space visibility testing framework used for SSAO can be extended to capture directional lighting and indirect light transport, not just simple occlusion.
  • Directional information is key for realism: Ignoring light direction leads to visually unconvincing results, especially with colored light sources. SSDO's ability to capture this directionality significantly enhances realism.
  • One bounce is a significant improvement: Even a single bounce of indirect light adds substantial visual richness, particularly in scenes with colored surfaces, by simulating color bleeding.
  • Image-space flexibility for dynamic scenes: The image-space approach is a critical enabler for real-time performance in dynamic scenes, as it decouples computation cost from geometric complexity.
  • Time-quality trade-offs are essential: The paper explicitly discusses how extensions like depth peeling and multiple cameras improve quality at the cost of performance, offering a practical framework for artists and developers.

Key Results

  • Performance: The method achieves real-time frame rates (e.g., 20.4 fps on a 537k polygon scene at 1600x1200). Adding SSDO incurs a modest overhead (+3.6% over SSAO), while adding one bounce increases it to +31.1%.
  • Visual Quality: The results show significant improvements over standard SSAO, with visible colored shadows and indirect light bounces. The paper claims visual equivalence to SSAO in the worst case without introducing new artifacts.
  • Dynamic Scenes: The method handles dynamic scenes and animated objects without temporal flickering.
  • Integration: The approach can be integrated with existing global illumination techniques like Instant Radiosity and environment map lighting to enhance their realism.

Limitations and Future Work

  • Bias: The screen-space approach is inherently an approximation and can suffer from sampling bias, especially for distant objects or when the scene structure is not well-represented in the framebuffer.
  • Limited Bounces: The current method only approximates one bounce of indirect light.
  • Approximated Form Factor: The calculation of indirect light uses an approximated form factor, which can lead to minor errors.
  • Future Work:
    • Investigating unbiased computation of indirect light in screen space.
  • Storing blocking information per pixel for temporal reuse.
  • Extending to multiple indirect bounces, specular materials, subsurface scattering, and importance sampling.
  • Exploring acceleration structures (e.g., hash tables) to replace framebuffer projection for proximity queries.
  • Combining screen-space and object-space approaches for optimal GI solutions.

Summary updated

BibTeX
@inproceedings{ritschel2009approximating,
  author = {Tobias Ritschel and Thorsten Grosch and Hans-Peter Seidel},
  title = {{Approximating dynamic global illumination in image space}},
  booktitle = {SI3D},
  year = {2009},
  doi = {10.1145/1507149.1507161},
}
Figure