-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA256
Hi Steinar,
On 10/22/2015 12:05 PM, Steinar Midtskogen wrote:
> 3.1% doesn't sound very unreasonable if this means you're applying
> the filter to all coded blocks without any signalling.
>
> Thor has a bit for every 64x64 block, except blocks that are
> entirely without coefficients (can still have vectors, though) or
> entirely bi-predicted. Such blocks are never filtered. If a 64x64
> block is filtered, subblocks can still be left unchanged if they
> lack coefficients or are bipred.
>
> The test whether to filter a 64x64 block or not is a simple
> squared difference with the original block. If the sum for the
> filtered block is less than the sum for the unfiltered block, the
> filtered block is kept.
Actually, the 3.1% is with signalling, in exactly the same way as we
did the signalling for deringing or for the first CLP filter. The test
is based on a PSNR-HVS-like metric compared to the original block. The
flag is entropy coded (with context from the neighbouring blocks),
with an RDO decision. Our superblock size is 32x32 so the cost is a
little bit higher than for Thor, but not significantly so since we
have entropy coding. We do not signal a flag on superblocks that are
entirely skipped. For superblocks that are filtered, we still disable
the filter in areas that are skipped. I believe this is pretty much
what Thor does too.
> It's interesting to see that the deringing also works reasonably
> well for low bitrates where CLP is falling apart (though I believe
> it still is useful subjectively).
>
> I did some experiments also signalling offsets larger than 1 (same
> offset for every 64x64 block) which probably helps more at low
> rates than at high rates. But even if I assumed zero signalling
> costs, the gains weren't big.
Indeed, for lower bitrates you need to be able to have larger offsets.
The key is being able to avoid blurring out edges, which is what our
deringing filter attempts to do.
> An observation in Thor is the CLP filter works better if applied
> after the deblocking filter and not before. I think SAO in H.265
> has a similar property.
Yes, both CLP and deringing in Daala are applied after our lapping
post-filter. We never tried applying it before, but I would not expect
anything good to come out of it.
Cheers,
Jean-Marc
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2
iQEcBAEBCAAGBQJWKQ1IAAoJEJ6/8sItn9q9oEgH/3gN2aaDF9BW7724NTNWHjoK
xsbvnvAD8GwiVm6VB7ryH34pWSle8Cm8bWJ2U58bjiTw0cU8sHvsV0HxaCsceNyf
Cu1fpasJal1mWW1uOtSfvsjP4dyl0prq/mciHwCyG5Q66e2oKEKCyxCOm4hvQKgM
guqa3jM178EC4WStdIdccFny6AHrp/N0qeweCayQuvYJHt6gw8wSucRiibEQNi3y
2b3NSk3npOJxO90B848KMKgwzu3Za+nqf/7sP56HQ3mydBhIFjsi3f1QdhpXFefl
kTQpm+Ll1d9WAHVM4I+JHXqwH38qbRMrCiAz2Nz+XDa2GMlm06nc03sC088keRQ=
=M9NR
-----END PGP SIGNATURE-----
_______________________________________________
video-codec mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/video-codec
