vp8 probabilty models.

[vijay]

Hi,

I see that in case of VP8 probabilty models are updated across the
frames unlike h264(limited within slice).
Is it done for better probabilty estimate?

Second I observed that we are sending full probabilties in frame
headers incase of different probility updates. Can't we send only
difference of probabilties for more compression. For simplicity
difference value can be limited to some 4 bits. Statiscally also
variation in updates shouldn't be much but It can give better
compression.

Thanks for reply in advance,
Vijay.

[Rajesh Pamula]

Hi,

I do not have the spec with me right now, but I remember that there is a 1 bit field which indicates whether the probability of the token is updated or not.

Sending differences might be a decent idea but i limits the swing in the probability values. If you were to consider the whole range then differences would only worsen the situation by adding 1 extra bit per token probability. Also added to it that is increase in arithmetic complexity, which of course can be neglected.

And VP8 probability update at frame level is probably done to simplify the entropy coder. It is likely to be inferior to that of AVC which uses CABAC.

Thanks and regards,

Rajesh

--
You received this message because you are subscribed to the Google Groups "WebM Discussion" group.
To post to this group, send email to webm-d...@webmproject.org.
To unsubscribe from this group, send email to webm-discuss...@webmproject.org.
For more options, visit this group at http://groups.google.com/a/webmproject.org/group/webm-discuss/?hl=en.

--
Rajesh Pamula
Post Graduate Student
Analogue and Digital Integrated Circuit Design
Imperial College London.
Residence Phone: +44-20-78522056

[vijay bansal]

Hi,

Thanks for the quick reply. I have seen in the encoder code that
arithmatic encoding is done on full frame in second pass. What I mean
is first it is doing pixel encoding=>mode/MV/token generations for
whole frame.So intermediate data is stored in intermediate buffers.
Then based on statistics collected while encoding in first pass it
updates the probabilty models. Then it does the encoding of full frame
in oneshot. But drawback of this scheme is more SDRAM accesses are
required due to full frame equivalent intermediate buffers.In bigger
resolutions these accesses would result in more power dissipiation. So
is there method todo on the fly encoding without intermediate
buffering. Using previous frame probabilty may result in loosing
coding efficiency.

Regards,
Vijay.