just a thought on quant matrices

[mNEMOsyne]

hi!
i just thought about the quantization matrices feature and since i like the quant modulation koepi introduced i thought forward:
what if we could let the codec modulate the quant matrix 'on the fly' for each frame depending on the motion and on the quantization level? so we could get a mpeg like quant for lm and a .h263 like quant for hm but not only this or that one, the quantization could be much more specific for each frame. maybe we could use even sharper quant for lm and quant level 1+2 than mpeg is and mpeg for medium motion / quant levels..

regards
mNEMO

edit: hope you got what the idea is ;)

[-h]

This has been considered, though in a different manner to what I think you're suggesting.

There is a significant overhead involved in sending the matrices - just look at how many numbers are in there :) Plus, you can only specify a custom matrix in a VOL header, and those are larger than the usual VOP headers which get stuck to each MPEG4 frame.

One of the ideas Isibaar was tinkering with, was to detect whether a block was high or low motion, and depending on that, "play" with the quantized values for that block so that they take up less space. By doing this, you eliminate the need for sending and switching between big matrices all the time.

I'm not sure what the status of such "encoder post-processing" is (-very- different to post-processing in the decoder, which everyone's used to), but progress will likely be made sooner or later. Don't you just love open source development? Ah, no deadlines :)

-h

[mNEMOsyne]

>One of the ideas Isibaar was tinkering with, was to detect whether a block was high or low motion, and depending on that, "play" with the quantized values for that block so that they take up less space. By doing this, you eliminate the need for sending and switching between big matrices all the time.<

so that would result in a feature like motion masking!?nice! how much is it implemented in the(upcoming)api yet?

mNEMO