On 17/03/2026 17:49, Christophe Gisquet wrote:
Hi,
Le mar. 17 mars 2026, 16:13, Lynne via ffmpeg-devel <ffmpeg-
[email protected] <mailto:[email protected]>> a écrit :
So I based this on the best lossy variant, which was Blackmagic RAW. It
does a partial debayering, but here, I keep the difference instead of
throwing it away. Then I use the regular FFv1 prediction scheme to
encode it.
The prior art for the decomposition is "Reversible color transform for
Bayer color filter array images" by Iwahashi et. al.
I imagine if complexity isn't super critical, but I have no idea about
parents (which are critical for ffv1?), you could run a stronger
decorrelation process. Maybe not PCA, but a simpler linear model (a la
CfL/LM) built from running stats, or neighbourhood.
But yes, for Bayer and/or RGB and/or higher-chroma-sampling images, that
sounds like a great idea.
Definitely, I did mention this in my follow-up to Michael. We could --
after all, custom RCT coefficients are still marked as unstable, so we
could rip out RCT with something much more aggressive. Vulkan would
definitely be able to handle much more aggressive searching.
Since it would influence coding RGB and Bayer, I think that this is
something that we should explore after we have implemented Bayer to have
a benchmark for both.
_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]
