adaptive custom quant matrices in theora

[bond]

just had an interesting discussion with derf explaining how matrices are handled in theora and it seems to be very powerful

theora allows to set own matrices for each quant, which practically means that you could use a low bitrate matrix (eg h263) on frames which get highly compressed (eg high motion) and a high bitrate matrix (eg hvs-best) on frames which have lots of details.
you could set a different matrix for each quant, but it would be also enough to set only one used for all quants (as possible in mpeg-4) or two for each quant 0 and 63 (in between the values would get scaled than)

any thoughts about this?


edit: of course as each matrix gets linked to a specific quant the matrices can even vary inside each frame as theora supports adaptive quantisation :D

[708145]

theora allows to set own matrices for each quant, which practically means that you could use a low bitrate matrix (eg h263) on frames which get highly compressed (eg high motion) and a high bitrate matrix (eg hvs-best) on frames which have lots of details.
you could set a different matrix for each quant, but it would be also enough to set only one used for all quants (as possible in mpeg-4) or two for each quant 0 and 63 (in between the values would get scaled than)

any thoughts about this?


I read about this in the theora specs about a year ago... very interesting read!

But as long as we don't see a codec making good use of it - well - it's pretty useless. :(

bis besser,
Tobias (who is generally disappointed by theora encoder quality)

[bond]

I read about this in the theora specs about a year ago... very interesting read!:)

But as long as we don't see a codec making good use of it - well - it's pretty useless. :(yep, derfs encoder uses it (defines matrices for 3 different quants afaik) but its more a proove of concept encoder it seems (doing some funky adaptive quantisation hvs things...)

[708145]

:)

yep, derfs encoder uses it (defines matrices for 3 different quants afaik) but its more a proove of concept encoder it seems (doing some funky adaptive quantisation hvs things...)

jep, I found out about this in another thread just after posting :P
Well... bad timing.

I'm really looking forward to a good theora implementation. And this encoder could be the one. :)

bis besser,
Tobias