Search found 50 matches

by Dietger
Mon Oct 08, 2012 1:43 pm
Forum: General Development
Topic: MLT Singularities
Replies: 7
Views: 4752

Re: MLT Singularities

Not that I know of. The best you can do is design a mutation strategy that does not get stuck in the first place (like Wenzel's manifold exploration for caustics). The better the mutation strategy is at finding high energy paths (the further away from the current path the better), the less correlate...
by Dietger
Sun Aug 19, 2012 6:56 pm
Forum: General Development
Topic: ERPT - fixed length mutation question
Replies: 6
Views: 4872

Re: ERPT - fixed length mutation question

:D couldn't agree more!

I secretly hoped to provoke some discussion with my previous post, but you said it all! ;) It's about time we see ERPT for what it really is...
by Dietger
Tue Aug 14, 2012 2:58 pm
Forum: General Development
Topic: Path Tracing coherence at Giga scales
Replies: 6
Views: 4622

Re: Path Tracing coherence at Giga scales

"Rendering Complex Scenes with Memory-Coherent Ray Tracing" by Matt Pharr comes to mind.
by Dietger
Tue Aug 14, 2012 2:49 pm
Forum: General Development
Topic: MIS on the GPU
Replies: 41
Views: 27265

Re: MIS on the GPU

There should not be any such black pixels. Are you sure you handle the whole shading normal vs. geometric normal correctly? Seeing that you use xpNs as input, it seem you use the shading normal to compute the probability per unit solid angle. If you don't correct for this later, you are bound to see...
by Dietger
Mon Aug 13, 2012 9:30 pm
Forum: General Development
Topic: MIS on the GPU
Replies: 41
Views: 27265

Re: MIS on the GPU

That is completely true when computing the probability for sampling a bidirectional path. However, the MIS weights sum over all bidirectional strategies, each having its connection at a different edge on the path. They all have their deterministic connection probability of 1 at a different point. If...
by Dietger
Mon Aug 13, 2012 6:44 pm
Forum: General Development
Topic: MIS on the GPU
Replies: 41
Views: 27265

Re: MIS on the GPU

Nope, you should not. Remember that the MIS accounts for all other sampling strategies as well, some of which sample y_s as part of the light path or z_t as part of the eye path. For P_A(y_s -> z_t) you should take the probability of sampling z_t from y_s as if z_t was part of the eye path. So basic...
by Dietger
Fri Aug 10, 2012 4:52 pm
Forum: General Development
Topic: ERPT - fixed length mutation question
Replies: 6
Views: 4872

Re: ERPT - fixed length mutation question

To see why ERPT with fixed chain lengths is unbiased, you should realize that ERPT with fixed chain lengths essentially reduces to Veach's two-stage MLT and thus is unbiased. As a reminder, veach's two-stage MLT looked like this: 1. Sample Np paths X_1 ... X_Np with intensities I(X_i). 2. Compute me...
by Dietger
Fri Aug 10, 2012 3:03 pm
Forum: General Development
Topic: Veach thesis - formula question
Replies: 86
Views: 48710

Re: Veach thesis - formula question

These are just yet other sampling strategies, so you should just mix them with the all the others bidirectional strategies using MIS. If an eye path hits a light source, you should accumulate the MIS corrected radiance, weather the last vertex was specular or not (this is already accounted for by MI...
by Dietger
Wed Aug 08, 2012 9:34 pm
Forum: General Development
Topic: MIS on the GPU
Replies: 41
Views: 27265

Re: MIS on the GPU

For the recursive MIS, you suggest to look at my tech report 'Recursive MIS Computation for Streaming BDPT on the GPU' http://graphics.tudelft.nl/~dietger/paper.pdf . It contains the same MIS formulation as in my thesis, but slightly more general (and with a more compact derivation). I use a constan...
by Dietger
Mon Jul 30, 2012 9:17 pm
Forum: General Development
Topic: MIS on the GPU
Replies: 41
Views: 27265

Re: MIS on the GPU

This way, you will not have the same number of path per depth ! true, but that is exactly the same for regular BDPT: There is only 1 bidirectional strategy for making a path of length 1, 2 strategies for length 2, etc. To see that this is the same as for BDPT, consider what happens in SBDPT if inst...