-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA256 Hi Steinar,
On 10/22/2015 04:48 AM, Steinar Midtskogen wrote:
> It looks like you're adjusting pixels by one if just one or more
> neighbour is more than 1 higher or lower. I wouldn't expect that
> to work. How does this work:
>
> y[i*ystride + j] = x + ((((x0 > x) + (x1 > x) + (x2 > x) + (x3 > x)
> > 2) - ((x0 < x) + (x1 < x) + (x2 < x) + (x3 < x) > 2)) << 4);
Indeed, I clearly saw the "> 2" in the draft... and then omitted it
from my code. Using your code above, I am now getting a 3.1%
improvement. Although it's still lower than what you report in the
draft, it's not so far considering that our test set does not include
the two clips that are giving you the highest gains. For example, I'm
seeing that CLP at high rate improves BQTerrace by about 8% and Johnny
by about 9%. In fact, at very high rate (> 0.2 bpp), the deringing
filter and the CLP give very similar improvement. Deringing is mostly
providing better gains compared to CLP when the bitrate drops around
0.05 to 0.1 bpp (depending on the clip).
I'm attaching a PDF with the PSNR and PSNR-HVS curves for CLP and
deringing compared to no filter.
> Thomas implemented a simplified SAO for Thor for reference, just
> edge offset and +/-1, but this gives most of the gains in H.265
> anyway. The objective gains in Thor were useful, but not very
> convincing. I think Thor needs less deringing than H.265. Perhaps
> it's because the interpolation filter is different (fewer taps, the
> special 2,2 filter and possibly an improved frequency response).
>
> I will look into your filter, but I might not be able to implement
> anything before the meeting.
Well, most of the Daala code we have for the deringing filter should
be directly applicable in Thor. I can help in adapting it too. The
main thing that would be required from the existing Thor codebase is
doing the buffering so that the deringing filter only gets applied on
non-filtered pixels rather than "refiltering" some pixels. This is
important because the filter is longer than for CLP, so we can't just
avoid filtering on the superblock boundaries. So let me know if/when
you're interesting in trying this.
Cheers,
Jean-Marc
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2
iQEcBAEBCAAGBQJWKQCPAAoJEJ6/8sItn9q9AAMH/i/732JHBpExb/GzyuD82wSv
DCY+l868eLi+1ShLMYvcCBqmlIrkLxu9Ve+54BM7JYFJXHTyjva0urtjpCbqVqKT
KlDBHNRenQzOWPTg3BpuhVfR2FYYgWVmW4mNy/40QZxzsIQh7tHG07gYk9BJa4hk
A6ccLhUgIXYKdt3XWb3otrQhuAdTqQLSSARcdhDzXMGxsq/l8ti/M/Pm4Ptq+Vux
WQGgGzzEnjjvRmQ+rWxhWt1gcee916MQmpPyx7+OZl7hlSrY0drukU17TYNzZfjp
qbbtavpZlZ7sTWtOFiBQ97dPNXM70rBhuUEZfuQghw4QTx7WQeNulH/qdEUPp7E=
=FeCq
-----END PGP SIGNATURE-----
clp_vs_deringing.pdf
Description: Adobe PDF document
_______________________________________________ video-codec mailing list [email protected] https://www.ietf.org/mailman/listinfo/video-codec
