[XviD-devel] Coefficient thresholding

[Christoph Lampert]

On Sun, 18 Aug 2002, Milos Spasovic wrote:

> Hi,
> 
> I folowed discussion about adding a coefficient thresholding and syskin
> saying that improves PSNR for 0.1 dB for fixed bitrate. That might be
> true and I trust syskin because I currently don't have access to my
> Linux machine which I use together with gruels excellent tool for
> measuring PSNR. However I have one trailer which I constantly use for
> checking different motion estimations with fixed quant 3 and MPEG
> quantization. So I tested it with coeff thresholding enabled and there
> IS a drop in filesize compared to older builds - about 600 kb. To get
> to the point, after watching this trailer I noticed that too much
> details were removed, which was the most noticable on human faces. It
> looks like TOOSMALL_LIMIT is set too high (or is that too low?). So I
> was wondering if this coeff thresholding is necessary because the main
> thing that makes XviD so special is IMO the amount of details it keeps.
> So please if someone else can also compare with and without tresholding
> I would be very grateful.

We could do the following: 

A TOOSMALL_LIMIT of 2 already removes the danger of zero-iDCT at quant=1.
At quant=2 and higher, zero-iDCT did not happen once in the tests I did
since yesterday (for quant=1 without TOOSMALL it was about 6 blocks
per 100 CIF-frames). (around 0.003% of all blocks for quant=1).

We make coefficient threshholding a flag. If it's switched off,
a value of 1 (no threshholding) is used for quant>=2 and a value of 2
is used for quant=1. I doubt that for quant=1 the effect will be
very visisble...

All problems "gone", but it still can be switched off. 

Christoph


-- 
Christoph H. Lampert chl@math,uni-bonn,de | Diese Signature wurde maschi-     
Beringstr. 6, Raum 14 Tel. (0228) 73-2948 | nell erstellt und bedarf
Sprechstunden: keine, aber meistens da    | keiner Unterschrift. AZ 27B-6