On Wed, Oct 21, 2015 at 05:07:30PM -0400, Jean-Marc Valin wrote: > -----BEGIN PGP SIGNED MESSAGE----- > Hash: SHA256 > > Hi Steinar, Thomas, > > I saw your new constrained low pass filter draft for Thor and decided > to start experimenting with it. Putting the filter in Daala, I saw a > PSNR bd-rate improvement around 2.1% in the 0.05 bpp to 0.1 bpp range > over our ntt-short1 test set (I disabled our deringing filter for this > comparison to avoid the interactions). However, I was unable to see > improvements of 4-7% like you are reporting in the draft. > > Of course, it's quite likely that I messed up the implementation or > the adaptation to Daala. It's also possible that somehow Daala cannot > benefit from the CLP as much as Thor does. This is the code I wrote > for filtering a small block of pixels: > https://pastebin.mozilla.org/8850049 > I've had to do some adaptation since the Daala pixels are shifted up > by 4 at the point where we're applying the filter. Does this look > correct to you? > > The other experiment that would be nice to try is using the Daala > deringing filter in Thor. In Daala, our deringing filter is improving > PSNR bd-rate by about a 6.5% in the 0.05 bpp to 0.1 bpp range. The > deringing filter is documented in this IETF draft: > https://tools.ietf.org/html/draft-valin-netvc-deringing-00 > or (for those who don't like ASCII equations) in this pdf: > http://jmvalin.ca/notes/dir_dering.pdf > I uploaded some still image examples at: > http://jmvalin.ca/video/deringing_examples/ > Any interest in trying the Daala deringing filter in Thor?
Looking at chapel subjectively, I think an option to use a weaker filter would be helpful. Generally, the filter gives good subjective results. But some of the details might get lost (e.g. the last steps leading to the chapel). _______________________________________________ video-codec mailing list [email protected] https://www.ietf.org/mailman/listinfo/video-codec
