[FFmpeg-devel] [PATCH] RoQ video encoder

Eric Lasota riot
Sat Jun 2 21:01:52 CEST 2007


Michael Niedermayer wrote:
> its hard to comment on code i dont see ...
> why not post the relevant part of the file?
http://icculus.org/~riot/roqencpatch_060207_1432.diff

There are my current changes to it.  Converting it back into 4:4:4 
hasn't been done yet.
* numCB4==256 forces numCB2=256 to comply with Q3 decoder
* RoQ_ID_MOT is now handled properly
* Possibility sorter replaced with two-step list builder

I'll see if I can get the 4:4:4 conversion done (I think that just 
requires changing the eval functions and the common apply code, which 
will also fix the decoder), and make reduce_p_lists operate on a target 
quality and only use pure CBR if it needs to cap the frame size to meet 
the max rate.  The pure CBR function might be useful for getting a good 
starting lambda too.

> also the lambda factor could be used when deciding the codebook size
I'm not sure this is even possible.  Codebook size is a product of which 
entries get rejected from the image from other optimizations, it isn't 
really controllable or predictable, which is why it's currently not 
considered at all when weighting options.

> and last, id like to repeat that pure CBR is a very bad idea
> you could look at lambda as a value which says how much distortion
> you would be willing to trade for 1 bit, so very wildly variing it between
> frames to maintain a constant size per frame is like saying one bit of this
> frame is worth X distortion while 1 bit in the next frame is worth 100*X
> distrortion ...
>
> also id say libavcodec/ratecontrol.c could be used if you want ratecontrol
> but this can easily be added later, that is after the encoder is in svn ...
ratecontrol.c is laced with MpegEncContext dependencies, so I'll see if 
I can learn something from it, but the rate control stuff is probably 
gonna have to be done specifically for this the same way the motion 
search was.

Again, I know nothing about this sort of stuff, so if you have any 
advice on an algorithm or approach to use, I'm listening.




More information about the ffmpeg-devel mailing list