Posted By
bubis
on 2018-10-31
08:35:05
|  Re: DFLIConv2
Hi Ingo,
I haven't tried to make it adaptive yet, but that is a good idea. Maybe even a pre-optimization phase would make a difference, some analisis to decide where you need more luma or more chroma badlines.
First, I want to make sure that I have an optimizer that produces very close results to DFLI mode in generic mode with DFLI badline settings. Those should be equivalent, but normal DFLI consistently produced less conversion errors until now. (I explaind the reasons for this above.)
"Until now", because last night I added a multi-variable optimization algorithm to the mix (it perturbs one to three variables at once if the overall number of combinations is below a limit) and now that sometimes produces less error for the generic case.  It tries all chroma-chroma, chroma-luma, luma-luma pairs and all chroma-luma-luma and chroma-chroma-luma triples even when those are (naturally) components of different colors. Most of the gain is coming from triple var optimizations.
Still, I couldn't come up with an image yet that produces much better results with some unconventional badline settings. Maybe adaptation is the key. It is hard to believe that DFLI is always the optimal badline sequence.
|
