Adaptive quantuzation (once more...)

[Kurosu]

Reading all the works made around lumi masking, I wonder if several algorithms I've studied a bit were considered. I'm asking it here, as xvid forums are closed...

have XviD developpers considered using MPEG-2 Test Model 5.0 or H.263 TMN8 adaptive quantization algorithms ?

Additionally, on rate control, was the MPEG-4 Validation Model used? It contains a quadratic model linking picture size (in bits) to quantizer in that way:
T = a/Q + b/Q²
a and b are determined for every image type and linearly regressed from the encode results of several previous frames of that type (that algorithm is very likely to have bad results).

I don't think such points were discussed, so I'm asking...

[-h]

I believe gruel is currently investigating adaptive quantization, but has chosen to look at current publications on the matter instead of implementations from the standards.

The Verification Model's rate control was not used. An RDO scheme would be best, but in its place Foxer wrote a leaky bucket system which tracks quant decision error over time, and it's working quite well.

-h

[MfA]

What XviD presently uses as far as adaptive quantization is concerned is closest to TM5 in that it only tries to guess a masking factor per MB, but TM5 in effect tries to use texture masking and XviD uses luminance masking instead. It's not very effective though (either way).

Personally I dont like the traditional approaches, wether they be based on masking or on RDO. For one thing a multipass approach which can take into account the dquant coding costs is plain better (with RDO this comes down to the trellis based approaches from people such as Dirk Farin) for another IMO the allowable local distortion at a given local rate should be under user control, and not a result from automated optimization of some global RD equation which has little perceptual relevance.

Marco

PS. -h are you still going to try to implement the p-space ratecontrol? :)

[Kurosu]

I believe gruel is currently investigating adaptive quantization, but has chosen to look at current publications on the matter instead of implementations from the standards.

Indeed, all the models I was speaking of are rather old now(1995-1996), but nowadays all is Rate-Distorsion Optimized... I've never gone deep enough that matter, but it seemed for many algorithms awfully slow.


The Verification Model's rate control was not used. An RDO scheme would be best, but in its place Foxer wrote a leaky bucket system which tracks quant decision error over time, and it's working quite well.

Indeed. I guess this is also quick and quite adapted to Variable BitRate (as it seems most people do that).

Thanks for those explainations, you satisfied my curiosity :)

[-h]

PS. -h are you still going to try to implement the p-space ratecontrol? :)

All in good time :)

It amazes me how easily I can get sidetracked.

-h