DaalaMeeting20140304

From XiphWiki
Jump to navigation Jump to search
# Meeting 2014-03-4

Mumble:  mf4.xiph.org:64738

# Agenda

Status of current patches waiting review and commit...
- roll back det1?
- JM's quant matrix stuff etc
- metrics (and CI)
- reviews
LibrePlanet

# Attending

xiphmont, gmaxwell

gmaxwell: working on the det1 DC basis problem; trying the easiest most obvious thing first (optimizing based solely on DC leakage) didn't work out so well.
gmaxwell: next tried coding cain + leakage, results not impressive
http://jmvalin.ca/video/signal_dc_nondet1.png
http://jmvalin.ca/video/signal_dc_det1.png vs https://people.xiph.org/~greg/railway.png
These are all DC only images with the last being the det optiimzation with it limited to low leakage results and using leakage as a weak weight in the search.
xiphmont: right now, given that scaling has 'taken the pressure' off of the need for det1, if we have to choose det1 vs the smoother filter, smoother wins.
gmaxwell: Right. Though there is a difference in trade-off for low vs high rate. E.g. at lossess the det1 is a big win— we can't do the scaling there. But I've also wondered if we're not wasting or time dorking again with the filter stuff at this point.
xiphmont: agreed
gmaxwell: Also, putting down and picking up the toolsets/states repeatedly can waste some time and effort swapping.  It makes some sense to keep chasing this for now, because we ideally want det1 that's smooth.
xiphmont: I've been working on visual comparisons of the recent changes, as we seem to have not been making them of all changes. 
gmaxwell: I have been comparing visually, though mostly at high rate.  There det1 appeared to be a wash and the metrics mostly agreed.
xiphmont: I compare at lower rates so things are easier to see/notice-- but that may not be giving fully representative results either
gmaxwell: I agree that the lower rates are way more useful. I have a somewhat bad habbit of looking at the higher rates because they're more useful for outright broken code. (The image should look perfect or the code is broken) but thats not a good approach when evaluating trade-offs. I'm making a mental note to avoid continuing to make the same mistake. Probably as practice we should make a habbit of doing both a low and a high rate comparison.
xiphmont: re, jm's 209 set of patches (quant matrix for PVQ) they're an obvious win on smooth backgrounds that would otherwise block out badly (eg, subset2 Abandoned Railway signal) and are even not bad for detail there.  OTOH, it's harder to tell if it's a good think on images with more sharp forground edges like Streptopelia.  Probably a win overall, need to look at this more without det1.
gmaxwell: perceptual theory says that quantization matrixes should be flat when all distortion is visible, so there is a bunch of rate sensitive tradeoffs with the quantization matrix stuff.
xiphmont: ah, right.  I should be looking at a range of rates.  And activity masking.  
gmaxwell: which got us big SSIM improvements in Theora.  In daala, activity masking looks better but does not improve SSIM.  This is a mystery we should discover, because it may indicate a bug.
xiphmont: BTW, shall we back out det1?
gmaxwell: jm wants det1 backed out, gmaxwell agrees, xiphmont agrees.  No vote yet from derf, unlord, jack.
gmaxwell: other subjects… what do we want to prep for libreplanet?
xiphmont: I was hoping to have one more demo by then, and was considering putting off demo5 and doing a demo update covering some of the recent activity including the det1 stuff. I would be interesting to talk about some of the recent failures.
gmaxwell: might be fun to have a realtime motion vector analysis display running for demonstration purposes.  Currently fast enough at low res (probably).
xiphmont: just do it on the GPU!
gmaxwell: die.
xiphmont: I tried to upload an updated patchset to issue 209 and rietveld nacked me for not being the owner.  Nathan said that was undesireable and should get fixed.... did he go off and fix it?
gmaxwell: dunno.