Re: Misc Codec Questions....

> Oh okay, so you compare an arithmetic coder with a simple order-0 or
> order-1 model to something with a model that better matches the data.

The results for using his method in actual DCT coding are only 3.7% 
better than Huffman coding. He cites a reference stating that merely 
replacing Huffman encoding by arithmetic encoding gives 5-10% 
improvements. It is also suspicious that he gives no results for 
arithmetic coding in this secion, unlike his other sections.

The wavelet scheme he presents achieves 0.5% worse performance than an 
arithmetic coder.

In both cases, his entropy coder cannot be inserted directly into any 
standard encoding schemes. If the first case, he rearranges the DCT 
coefficients so that all coefficients from the same subband are 
adjacent, which kills all the benefits of a block transform. In the 
second, he uses a scheme called "Morphological Representation of Wavelet 
Data", which I had never heard of before, instead of EZT or SPHIT.

This is what strikes me as the most limiting aspect of his coder: it is 
not very flexible. The underlying design assumes that any combination of 
symbols with the same sum is equally probable. Except for in the case of 
a stationary encoder with binary symbols, where the sum is the number of 
ones, and thus with the length represents the probability of a one, I 
don't have an intuitive grasp of how that reflects the source 
statistics. With a normal entropy encoder, you have lots of options for 
coming up with a probability distribution for the next symbol. For his 
coder, it seems the only thing you can do is rearrange the input, and 
hope you get locally stationary pieces.


--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to '[email protected]'
containing only the word 'unsubscribe' in the body.  No subject is needed.
Unsubscribe messages sent to the list will be ignored/filtered.