theora dev - Nov 2009 - Theora Fast Quality Reduction Transcoding

Hi.

I have been working on a tool whose goal is to reduce the bit rate of theora
video by decoding to DCT coefficients, reducing the entropy of the
coefficients, and re-tokenizing the stream.

I have successfully used the decoder source to extract the DCT coefficients
for each block, and I am able to capture any and all relevant information
about where the block of coefficients falls in the image, the frag list, the
MCU and so on. That is to say the entire decoder state for a given block of
coefficients is available to me in a callback.

I had thought at this point to use the tokenizer directly to construct a new
token stream. However the complexity of that module alone is daunting. It
relies on state information retained and perhaps also generated by the
encoder. I think the retained information like the huff tables and maybe
even lambda can be constructed easily enough. But I am concerned about the
possible use of generated data since we are not doing a full encode, and
that there may be traps waiting to be sprung on me.

I have also considered running a full encoder but without any real image
data being fed to it. The processing required to generate coefficients would
be replaced with a callback to retrieve coefficients generated by the
decoder. However, this approach is very likely to be unworkable as, at the
very least, the inter frame motion vectors and quite probably other useful
information would be missing without further intervention from me.

So I am still inclined to use the tokenizer directly on the modified
coefficient stream. I would be very grateful if anyone knowledgeable about
the tokenizing process or the encode process in general could offer advice
or warnings about either approach I described.

Thanks in advance,
Chris

Chris Cooksey wrote:
> I had thought at this point to use the tokenizer directly to construct a
new
> token stream. However the complexity of that module alone is daunting. It
You may be interested to know about rehuff, an experimental tool for
changing the codebooks in a Theora stream. The tool itself is here:
http://svn.xiph.org/trunk/theora-exp/
with examples/rehuff.c being the main driver and lib/recode.c doing most
of the actual work.

This is probably too simplistic for you, since it cannot, for example,
change a block from coded to uncoded if you quantize all of its
coefficients to 0 (assuming it does not require a MV), and does not
break the decoded coefficients out by the block they originally came
from. But it may be a simpler starting point than the full encoder.
> I have also considered running a full encoder but without any real image
> data being fed to it. The processing required to generate coefficients
would
> be replaced with a callback to retrieve coefficients generated by the
The problem here is that it would also have to choose the same MB modes,
MVs, qi values, and coded block flags for the coefficients from the
original stream to make any sense. These decisions are made with
reference to the original input frame, which you do not have. I agree
with your assessment that this approach is probably unworkable.

Hi Timothy.

Thanks for the information. It confirms most of what I have found.
> You may be interested to know about rehuff, an experimental tool for
> changing the codebooks in a Theora stream. The tool itself is here:
I do know about the rehuff tool. It was the foundation for what I have done
so far, which includes reporting each coefficient block and the surrounding
decoder state. I also updated it to Theora 1.1.

As for trying to encode without actual image data, yes that is a non-starter
it looks like.

I am still holding out hope for using the tokenizer directly on the modified
coefficients. But you did identify the major area of concern I have: if the
coefficients of an inter frame block all go to zero, I will need to update
the coded block list. For that matter, what if the entire inter frame
becomes uncoded? This greatly complicates what would otherwise be a
seemingly simple task.

The rehuff tool works by simply duplicating the entire frame header and then
emitting a recoded huff stream. This won't work if I need to change the
coded block list, which ideally I would like to if possible. It is here that
I am kind of stuck right now. My understanding of that part of the spec is
limited. Any hints regarding that part of the process as its implemented in
theoralib would be helpful, i.e. what are the relevant structures, etc. I
will be working through it myself to try to better understand it.

Thanks,
Chris.