View Single Post
Old 12th January 2017, 05:02   #4607  |  Link
burfadel
Registered User
 
Join Date: Aug 2006
Posts: 2,229
I found the same thing. I'm currently encoding so won't do another comparison, but from the testing I did a while back STAR was beneficial over UMH in quality and the speed difference was very minimal. You almost get Full search quality at an encode speed similar to UMH. In theory you would expect SEA to be better than STAR. I also did some tests comparing the subpel refinement. Basically for the clip tested 2 was the ideal choice.

Level of RDO is an area for improvement I think. I use 4 because it has a good speed/quality tradeoff (yes I know 4 is basically the same as 3 currently). RD level 5 produces nice results but there is a massive performance hit. Combine --rd 5 with --rd-refine and --opt-cu-delta-qp and it produces a nice result. --rd-refine and --opt-cu-delta-qp are only available in RD 5 and 6.

I'm assuming it's possible to shortcut the extra calculations in --rd 5 as has been done with TU intra and enabled with --limit-tu 3 (for example), and like with --limit-modes and using --amp (which then has a negligible performance impact). I think if something along those lines could be done with --rd 5 and made the new -rd 4, it would be great. Likewise, using --rd-refine and --opt-cu-delta-qp as default options in this mode. The --rd-refine I'm assuming could also be shortcut (and use the shortcut from the rd level). By shortcut I mean like --limit-modes, --limit-refs etc.

I believe this new --rd 4 level if it could be achieved with minimal performance penalty over --rd 3 would be great!

Also a much more mild Sample Adaptive Offset loop filter would be good because currently (as discussed previously by others and myself) it loses detail with no real benefit. I can see in theory it should be better to have enabled, I think having it at a strength of say, 30 percent of what it currently is would be good. Hey, why not give the option to specify a strength, with 1.0 being the current strength and 0.3 (or whatever is found to be the most subjectively appealing) being the default strength?

Another idea is --strong-intra-smoothing. The help states 'This flag performs bi-linear interpolation of the corner reference samples for a strong smoothing effect.'. How about bicubic interpolation?

Last edited by burfadel; 12th January 2017 at 05:11.
burfadel is offline   Reply With Quote