AOmedia next (AV2) codec

[BlueSwordM]

The main trick with CWG, aka proposals, is that they have to be rock solid to find themselves into production.

Their definition of rock-solid depends on a lot of members: hardware teams, software teams and worst of all, PSNR/SSIM scores.

Some of us have had to do a lot of work convincing peeps that a feature that increases perceptual quality and SSIMU2/butteraugli-jxl/XPSNR scores at the cost of non perceptual blur metrics, PSNR/SSIM.

2 features were added with the behaviors cited above and yet, some people complained :P

They did manage to go through, but it was hard.

[Jamaika]

Originally Posted by BlueSwordM

The main trick with CWG, aka proposals, is that they have to be rock solid to find themselves into production.
Their definition of rock-solid depends on a lot of members: hardware teams, software teams and worst of all, PSNR/SSIM scores.
Some of us have had to do a lot of work convincing peeps that a feature that increases perceptual quality and SSIMU2/butteraugli-jxl/XPSNR scores at the cost of non perceptual blur metrics, PSNR/SSIM.
2 features were added with the behaviors cited above and yet, some people complained :P
They did manage to go through, but it was hard.

I created something but not everything works for me under gcc.
I deleted CONFIG_ENABLE_IBC_NAT, CONFIG_IBP_WEIGHT.
Frames-I for AV2 work for me under libwebp2. I'll add plus for AV2 for large size 5212x3468 support.
av2enc_avx.exe -q 100 -444 -size 280x420 -effort 7 -threads 4 -pass 1 -tune ssim image_jpeg.wp2 -d output.av2
tune butteraugli workes with latest jpegxl. Delete default BT709:
/*if (img->mc != 0 && img->mc != AOM_CICP_MC_BT_709 &&
img->mc != AOM_CICP_MC_BT_601 && img->mc != AOM_CICP_MC_BT_470_B_G) {
ERROR(
"Only BT.709 and BT.601 matrix coefficients supported in "
"tune=butteraugli mode. Identity matrix is treated as BT.601.");
}*/

av1enc_avx.exe -q 100 -444 -size 280x420 -effort 7 -threads 4 -tune butteraugli image_jpeg.wp2 -d output.av1

I wasn't able to create a movie.
https://gitlab.com/AOMediaCodec/avm/...75decc71064bcd
ffmpeg_avx2.exe -i "video.mp4" -f rawvideo -vf scale=1920:1080: in_range=limited: out_range=limited -frames:v 500 -pix_fmt yuv420p - | aomenc.exe --verbose --limit=500 --input-bit-depth=8 --i420 --width=1920 --height=1080 --ivf --good --passes=1 --pass=1 --bit-depth=8 --fps=30000/1001 --threads=20 --cpu-used=9 --end-usage=cq --qp=0 -o avm_yuv420p10le.ivf -

Assertion failed: is_inter_compound_mode(mode), file c:\gcc1150\x86_64-w64-mingw32\include\av1\common\blockd.h, line 178

https://www.sendspace.com/file/gh11if

[benwaggoner]

Originally Posted by BlueSwordM

The main trick with CWG, aka proposals, is that they have to be rock solid to find themselves into production.

Their definition of rock-solid depends on a lot of members: hardware teams, software teams and worst of all, PSNR/SSIM scores.

Some of us have had to do a lot of work convincing peeps that a feature that increases perceptual quality and SSIMU2/butteraugli-jxl/XPSNR scores at the cost of non perceptual blur metrics, PSNR/SSIM.

2 features were added with the behaviors cited above and yet, some people complained :P

They did manage to go through, but it was hard.

Sir, you did the work of Heroes!


People treat PSNR like some sort of fundamental definition of distortion rather than the decades-old "eh, easy to calculate and better than the other obvious alternatives" metric it truly is. And the square root of squares gets you a free absolute value conversion.

Lots of folks with a computer science background but without a neuropsychology or vision science background can assume that PSNR is what's before psychovisual optimization. But it's ALL psychovisual optimizations. It's just some of them are so old we don't really think about how they came to be.

sRGB/Rec. 709 is a perceptual optimization that makes 8-bit encoding work well within the range of what good CRTs used to be capable of reproducing.

Gamma is a psychovisaul optimization based on a first-order approximation of the human visual and CRT EOTF, easily implemented in simple analog circuitry. PQ is the second order approximation (but assumes a specific fixed ambient light). The non-psychovisual implementation would be linear light values.

Quant/Lambda tables are psychovisual optimizations based on our more accurate perception of horizontal and vertical details versus diagonal. And historically the PAL and NTSC versions embedded the different default sample aspect ratios (720x480 versus 720x576). The non-psychovisual implementation would be a flat quantizer.

Frequency transforms like DCT and iDCT themselves are perceptual optimizations about how our eyes and brains detect edges more than absolute brightness and color spatially. The non-psychovisual implementation would probably be pixel-based. And square pixels themselves are more of a computational friendly approach. Why not triangles or hexagons?