(L) [2010/05/29] [Dietepiet] [Coherent Metropolis Light Transport] Wayback!Hi All,
I recently stumbled upon the paper "Coherent Metropolis Light Transport with Multiple-Try Mutations" and I was wondering if anyone tried to implement it? As far as I can see, compared to a single perturb mutation in standard MLT, MLTMTM requires something like 32 coherent sub-mutations per mutation, still resulting in a single contribution to the image plane. No doubt the number of rays/sec will increase, but I don't see why this is going to help convergence speed. Could someone explain that?
Thanks,
Dietger
(L) [2010/05/29] [_bouliiii] [Coherent Metropolis Light Transport] Wayback!Hello,
the trick is actually to use the expected value to accumulate "contributions" for all paths. So:
- you generate "candidates" and "competitors" paths
- you accumulate all their contributions
- you pick the "best" candidate and you continue
Sorry, the paper quality is rather low but I did not spend a lot of time on it at that time.
Cheers,
Ben
(L) [2010/06/02] [Dietepiet] [Coherent Metropolis Light Transport] Wayback!Thanks, that explains a lot! I missed the part where the contribution of all candidates and competitors was accumulated. It seems like an interesting method, but the paper is indeed a little summary. It would be nice to see some convergence comparisons w.r.p. standard metropolis. I'm going to re-read the paper some more and hopefully try it on an ERPT.
(L) [2010/06/02] [Dietepiet] [Coherent Metropolis Light Transport] Wayback!I still have some difficulty understanding the final accumulation of color on the image plane. Am I correct to assume that the contribution of the "candidates" and "competitors" paths is computed using the usual detailed balance for MPT and not the modified balance for MTMPT? It seems to me that, as MTMPT satisfies the usual detailed balance, this may be used for the estimation of path contributions.