Is it okay to ignore TransferCharacteristics and MatrixCoefficients? about libavif

[joedrago]

My YUV<->RGB conversion is very likely incomplete on some of those MatrixCoefficients cases, and in the subset that depend on the TransferCharacteristics, they might also be wrong. That said, I believe most of the most commonly used paths are correct, and I think typical YUV<->RGB conversion does not temporarily linearize with the transfer function.

I'll mark this issue as "help wanted" in case someone wants to contribute solutions to some of the more interesting combinations. Long term, I'd prefer libyuv to support all of these combinations (including 12bpc) and I can replace reformat.c entirely with that library.

[ledyba-z]

Thank you for replying!
I will survey about other project. And if I can find some information or solutions, I will share them.

Long term, I'd prefer libyuv to support all of these combinations (including 12bpc) and I can replace reformat.c entirely with that library.

Yeah, I also really want to replace YUV <-> RGB conversion functions with libyuv also in our implementations...

[joedrago]

As a note, you are never required to use libavif's YUV<->RGB ever, they are just provided for convenience. libavif can be used entirely with just the YUV planes in both directions, and you can convert to RGB completely independent of libavif.

[ledyba-z]

Thanks for notice!

Yeah, we use Apple's vImage Frameworks for iPhones (of course, they are only available in iOS or macOS).

[ledyba-z]

I found that H.273 (refered by AV1-spec) specify about conversion.

In a nutshell, just MatrixCoefficients affects a conversion matrix.

[joedrago]

With your recent accepted PR, I'm going to close this. Thanks for the code.