Sequential Monte Carlo instant radiosity

Peter Hedman, Tero Karras, Jaakko Lehtinen

Abstract

Instant Radiosity and its derivatives are interactive methods for efficiently estimating global (indirect) illumination. They represent the last indirect bounce of illumination before the camera as the composite radiance field emitted by a set of virtual point light sources (VPLs). In complex scenes, current algorithms suffer from a difficult combination of two issues: it remains a challenge to distribute VPLs in a manner that simultaneously gives a high-quality indirect illumination solution for each frame, and does so in a temporally coherent manner. We address both issues by building, and maintaining over time, an adaptive and temporally coherent distribution of VPLs in locations where they bring indirect light to the image. We introduce a novel heuristic sampling method that strives to only move as few of the VPLs between frames as possible. The result is, to the best of our knowledge, the first interactive global illumination algorithm that works in complex, highly-occluded scenes, suffers little from temporal flickering, supports moving cameras and light sources, and is output-sensitive in the sense that it places VPLs in locations that matter most to the final result.

Introduction: Solving the Flicker in Global Illumination

Simulating global illumination—how light bounces off surfaces to illuminate areas not directly hit by a light source—is a cornerstone of realistic rendering. A popular real-time approximation is Instant Radiosity, which simulates bounced light by placing Virtual Point Lights (VPLs) on surfaces throughout the scene.

However, in complex, highly-occluded environments like architectural interiors or dense game levels, most light paths never reach the camera. Blindly distributing VPLs wastes computation on unseen areas. Conversely, "view-adaptive" methods that focus VPLs only on visible areas often suffer from severe temporal instability. As the camera moves, the VPL distribution shifts drastically, causing jarring flickering and popping in the indirect lighting. This paper introduces a real-time algorithm that achieves high-quality, multi-bounce indirect lighting with unprecedented temporal stability by treating VPLs as particles in a Sequential Monte Carlo filter.

Methodology: Evolution over Revolution

The core innovation lies in framing VPL placement as an evolving probability distribution. Instead of regenerating the entire set of VPLs every frame, the algorithm carries the vast majority over from the previous frame, updating only a tiny fraction (~1%) to adapt to changes.

The SMC Pipeline

  1. Indirect Visibility Filtering: GPU ray casting ensures VPLs are placed only on surfaces capable of bouncing light to the camera.
  2. k-Nearest Neighbor (kNN) Density Estimation: Since VPLs persist across frames, their mathematical distribution becomes complex. The authors use a 3D kNN search to estimate local VPL density on the fly.
  3. Sequential Updating: The algorithm identifies "undersampled" regions—areas that are bright but lack sufficient VPLs. It discards a small, fixed number of the least useful VPLs (e.g., 16 out of 2048) and replaces them with new ones in these critical areas.
  4. Rendering: The final scene is rendered using standard interleaved sampling and shadow maps to compute illumination from the VPL set.

Key Results and Evaluation

The method was evaluated on challenging 3D scenes including Epic Citadel, Soda Hall, and a custom Maze environment designed to test severe occlusion.

Comparison of VPL distribution and temporal error across different methods. The proposed method (d) shows significantly lower temporal error (top row) and more focused VPL distribution (bottom row) compared to baselines (a-c).

Performance and Stability

  • Temporal Stability: The method dramatically reduces flickering. In the dynamic Epic Citadel scene, the temporal error metric was 0.0068, significantly lower than Metropolis Instant Radiosity (0.0219) and standard Instant Radiosity (0.0325).
  • Image Quality: Spatial quality matches high-end offline methods like MIR, but without the associated temporal noise.
  • Real-time Performance: On an NVIDIA GTX Titan X, frame times ranged from 33.2 ms to 47.2 ms (approx. 20-30 FPS) for 2048 VPLs.

Conclusion and Future Directions

By prioritizing temporal coherence through incremental updates, this research provides a robust solution for interactive global illumination in complex scenes. While currently limited to diffuse surfaces and facing challenges with dynamic geometry due to shadow map dependencies, the approach sets a new standard for stable, view-adaptive VPL distribution. Future work may leverage hardware-accelerated ray tracing to replace shadow maps, potentially resolving current limitations with moving geometry and non-diffuse materials.

Summary updated

Abstract

Instant Radiosity and its derivatives are interactive methods for efficiently estimating global (indirect) illumination. They represent the last indirect bounce of illumination before the camera as the composite radiance field emitted by a set of virtual point light sources (VPLs). In complex scenes, current algorithms suffer from a difficult combination of two issues: it remains a challenge to distribute VPLs in a manner that simultaneously gives a high-quality indirect illumination solution for each frame, and does so in a temporally coherent manner. We address both issues by building, and maintaining over time, an adaptive and temporally coherent distribution of VPLs in locations where they bring indirect light to the image. We introduce a novel heuristic sampling method that strives to only move as few of the VPLs between frames as possible. The result is, to the best of our knowledge, the first interactive global illumination algorithm that works in complex, highly-occluded scenes, suffers little from temporal flickering, supports moving cameras and light sources, and is output-sensitive in the sense that it places VPLs in locations that matter most to the final result.
BibTeX
@inproceedings{hedman2016sequential,
  author = {Peter Hedman and Tero Karras and Jaakko Lehtinen},
  title = {{Sequential Monte Carlo instant radiosity}},
  booktitle = {I3D},
  year = {2016},
  url = {https://openreview.net/forum?id=IDQQxdh0NIz},
}
Figure