[daala] Early results with new PVQ code
jmvalin at mozilla.com
Tue Dec 17 15:01:45 PST 2013
On 17/12/13 05:08 PM, Jack Moffitt wrote:
> It seems to zigzag less. Is that significant somehow?
The zigzags are caused by the scalar quantizer when using an odd
quantization step size. I don't think we ever got an explanation for it,
but PVQ does not have that behaviour. The little zigzaging left is dur
to the remaining scalar quantization because PVQ isn't implemented for
4x4 and 16x16 (even when forcing 8x8, chroma ends up as 4x4 scalar).
> On Tue, Dec 17, 2013 at 3:00 PM, Jean-Marc Valin <jmvalin at jmvalin.ca> wrote:
>> Actually, these curves may be more useful. They force all blocks to be
>> 8x8 (the only size supported with PVQ for now) rather than have block
>> On 12/17/2013 04:04 PM, Jean-Marc Valin wrote:
>>> Here are some early results I got on subset1 with the new PVQ code I've
>>> been working on. It can be enabled simply by changing:
>>> mbctx.run_pvq[pli] = 0;
>>> mbctx.run_pvq[pli] = 1;
>>> in encode.c.
>>> It seems to be performing very slightly better than scalar, which is a
>>> good sign at this point considering it just started working at all.
>>> daala mailing list
>>> daala at xiph.org
>> daala mailing list
>> daala at xiph.org
More information about the daala