In the second half of my time as a UC Berkeley student, I co-authored two papers in ACM. We created a fast image-space filter to reduce noise from ray traced diffuse interreflection (causes color bleeding) and soft shadows.
In ray tracing, a rendering technique to produce an image of a 3d scene, you can control the sample rate, or, how many times to estimate the value of the color of a pixel.
Certain effects require many samples to give an accurate pixel color. If these effects weren’t given enough samples, the result would be a noisy image. Each sample costs about the same as other samples, so the render times of a ray traced image grows roughly linearly with the number of samples you use.
In my last two years at Berkeley, I, along with Soham Mehta, Ravi Ramamoorthi, and Fredo Durand explored soft shadows and diffuse interreflection, two of the effects that typically require many samples.
We created filter that applies on a noisy rendered image (with some additional knowledge collected during rendering) to reduce the noise of these effects. This filter essentially shares samples between pixels when it is appropriate to do so (i.e. in regions that have very soft shadows).
The filter we created is based on the physics of the frequency and distribution of light. This means that our filter converges to the correct result, rather than being a heuristic that does not match reality. We also tell you the regions that require more samples, so you can sample those regions more.
I give a more in-depth look at the filter itself in my Master’s Thesis.
The derivation of the technique for soft shadows and diffuse interreflection can be found in our SIGGRAPH Asia 2012 paper, and SIGGRAPH 2013 paper, respectively.