Possible new way of compression?

[crusty]

I was just thinking about artifacts in dark scenes etc., and suddenly remembered that the human eye is much less sensitive for color in low-light situations.

Is it possible somehow to use this to add more compression to dark scenes?

I was thinking along the lines of this:
In a dark scene, reducing the precision of color matches, and allowing more error in color information IF that reduces bits in a macroblock to a certain degree.

After all, the darker the scene, the less precise the human eye can detect small colour difference. If this could be made into a feature that would be very interesting.

I know we already have adaptive quantization, but AFAIK this only works in the luma field (correct me if I'm wrong). This extra compression would also work in the chroma field.
And also AFAIK, the human eye is still very sensitive to luma changes, even at low light conditions. I guess that's why earlier versions of AQ were known to cause artifacts, for they were not properly tweaked for this. But my guess is a similar technique for chroma changes could be less intrusive.

Could one of the devs tell me:
1: If this would be possible
2: If this would be a good idea
3: If it's already done
4: If I'm talking out of my ass. :) (no harm done then)

???

[Teegedeck]

I think I remember there already were many ideas how to compress colour-information better. But sadly, colour is already stored at half resolution and takes up little space, so it may not be worth the effort.

[sysKin]

Originally posted by crusty
1: If this would be possible
(...)
3: If it's already doneNot yet, as adaptive quantization works on whole macroblocks (luma+chroma). It will be possible when I implement my system of HVS plugins. In fact, it will be very easy then.
2: If this would be a good idea
(...)
4: If I'm talking out of my ass. :) (no harm done then)Yes kinda ;p You said you're seeing artifacts in dark areas, yet you want to compress them even more. Kinda weird ;)

:)

[crusty]

I meant that Lumimasking could create artifacts in older alpha builds...not that it's broken now... :)

[scharfis_brain]

how about something like this schemeinto the encoding - decoding process:

video -> coloryuv(autogain=true) -> encoding -> decoding -> invert the autogain

this should solve problems in dark scenes (but not on dark parts of a bright picture)

[MfA]

It will also explode bitrate.

There are a million non standard compliant improvements you can come up with ... but the moment you do that you might as well forget about MPEG-4 altogether.