Ambient Occlusion
Introduction
This project's goal was to produce an implementation of Ambient Occlusion mapping which used graphics hardware to accelerate the occlusion calculation.

Ambient Occlusion is a technique developed by ILM during the making of the Pearl Harbor movie to improve lighting without requiring full global illumination computations. The algorithm is able to give the effect of more complex lighting and shadowing computations without the computations. (Obviously it is not as correct as the more complex solutions, but at times this is sufficient.)

Ambient Occlusion Algorithm
The Ambient Occlusion algorithm captures a general visibility term for the surfaces/vertices of an object. This visibility term approximates how much ambient light would arrive at the surface. In an algorithmic sense, this term is computed by casting N rays from each surface in the model and detecting how many of these rays intersect other geometry in the model and what rays don't intersect anything. For each ray that doesn't intersect any geometry, a cosine weighting is computed for the ray and the value is added to the visibility term for that surface. (A cosine weighting function is used if a uniform distribution of rays is used.) This process is continued for each surface until the entire model has been computed. A single floating point value is the result for each surface or vertex.

The images to the right render only the ambient occlusion term. There is no other lighting in the images.


42 viewpoints

162 viewpoints

642 viewpoints
42 viewpoints

162 viewpoints

642 viewpoints
Implementation
The above algorithm is the general description of the algorithm but not the implementation that I chose to do. Instead, I chose to use the concept of "Depth-Map Ambient Occlusion". The reason for using the depth-map based ambient occlusion is that using hardware we are able to compute the ambient terms much quicker than the ray tracing algorithm.

Depth-Map Based Ambient Occlusion
Depth-Map Ambient Occlusion approaches the problem from a different perspective. It renders N viewpoints from outside the model and checks to see which vertices are visible in the viewpoint, using a depth map, and then computes the visibility contribution for that viewpoint for those vertices.

To generate viewpoints for this implementation, I chose to use a icosahedron, and subdivided it until it reach the number of vertices that were desired for the processing. This accounts for the strange number of viewpoints used for each of the images above. The image below shows the viewpoints around the model.

From each of these viewpoints, the model was rendered into a depth map in the graphics board memory then the depth map was read back into system memory. Each vertex was projected into the depth map and its visibility was determined. If it was visible, the cosine weighting was computed and added into the ambient occlusion term.

Once all the views are computed, the model can be rendered with the ambient occlusion term as its color information. Obviously this is not the final use for the term, it should be used as part of the final equation for lighting the surface.

A bent normal can be computed to allow sampling environment maps to allow a nice complex ambient effect. (A bent normal is just the normal along the direction where the surface obtained most of its ambient light.)

No matter what the final use, the model can be manipulated in real-time with this level of lighting information.

References

  • Landis, Hayden - "Production Ready Global Illumination", Siggraph 2002
  • Whitehurst, Andrew - "Depth Map Ambient Occlusion"
glenn@raudins.com