To those interested: Trellis tests

[mf]

I just did a test with Trellis quantization. Note that this feature is not available in public builds, so if you want to test it yourself, build your own XviD build or wait until a public build with trellis comes out.

I did a test on anime, (2pass with b-frames) and these were the results:
Filesize with trellis is 4KB lower.
Min PSNR difference: 0,2065
Avg PSNR difference: 0,01859
Max PSNR difference: 0,0

These values are positive, meaning that using trellis at these settings raises PSNR, and thus increases quality.

This is an unorthodox way of PSNR measuring (usually you use constant quant), but in my opinion this gives a better idea of what to expect with real encodes.

If you have also tested trellis, please reply with your results. :)

[CruNcher]

hmm i got a little other result but im not sure if i calculated everything correct or maybe it has to do with our different sources

test clip = 1501 frames @ 25 fps 512x384


Devapi3 No Trellis Quant 2 = 12,6 MB (13.290.496 bytes) PSNR Min = 44.9026 Avg =46.0966 Max =47.2174 (Sum)* = 69190,9966
Devapi3 Trellis Quant 2 = 11,9 MB (12.543.488 bytes) PSNR Min = 44.9708 Avg =46.0827 Max =47.1961 (Sum) = 69124,05
Devapi4 Trellis Quant 2 = 12,6 MB (13.265.920 bytes) PSNR Min = 45.0427 Avg =46.1816 Max =47.3873 (Sum) = 69272,4


Sum = Average PSNR*1501 frames

so my conclusion is that Trellis R-D in Devapi3 has a slightly lower PSNR but acceptable for the Size win and if you compare Devapi4 Size and Quality with Trellis against Devapi3 without you can clearly see the Quality win and the lower size. But as i said im not sure if im correct here.

[Selur]

just found a little article about "Trellis Quantization":
Analysis of Trellis Quantization for Near-Lossless Image Coding (http://www.informatik.uni-freiburg.de/tr/complete_g.html)

might be interessting,..

[Lobuz]

There is at umaniac's site new version with commited R-D code. I've made a little test with q2 and compared to latest koepi's version. The size was a little bigger but I didn't measure quality so maybe it's better.

Regards
Lobuz

[Koepi]

As stated in hubhub's thread: VHQ R-D code is different from trellis quant (which works on the DCT of the macroblock and sets all coefficients to zero in cases the decision says so).

SysKin did another approach, he did a refinement of his BITS saving VHQ code (i.e. mode decision) based upon rate-distortion, too. The advantage is that PSNR (the "mathematical" quality) won't decrease but can only raise.

Regards
Koepi

[Koepi]

Please discuss these features at the "xvid-14052003-1.exe"-thread so we don't have to follow 3 different threads just to gain al the information. Thanks :)

Best regards
Koepi