10-14-10 - Image Comparison Part 5 - RAD VideoTest

VideoTest is my test video coder for RAD. It's based on "NewDCT" , in fact it has exactly the same DCT core, but it has a sightly better perceptual tuning, and it has a better RDO encoder.

log rmse :

scielab ms-ssim :

videotest vs. newdct is almost identical in rmse, but we did make a nice step up in perceptual measure.

I am finally beating JPEG ari in the perceptual measure, but it's a bit disturbing how much work I had to do! And of course PAQ JPEG still dominates.

The videotest I frame coder has pretty sophisticated RDO, but it's missing a lot of other things that modern coders have, it has no I predictors, no in-frame matches. It uses a little bit of a perceptual D measure for RDO, but not as well tweaked as x264 by a long shot.

videotest currently crashes for high bit rates; I remember I put some stupid fixed size buffer somewhere to get things working one day, and now I forget where it is :( So that's why there are no results for it above 2.0 bpp