Re: Advanced Motion Compensation for sub 500kbit DVD Video

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Saturday 20 July 2002 12:08, Johan Hanson wrote:
> Lourens Veen wrote:
> > A slightly different idea would be to do away with the concept
> > of keyframes, and instead use separate textures that are
> > streamed alongside the motion data. This would require a lot of
> > memory for texture caching on the decoder side though, [...]
>
> You don't *need* to encode textures as complete frames.
> You *could* cut and paste in the wavelet data, to send regions of
> a keyframe ahead of time.
> However, compression would be suffering because of lack of
> locality in the data and it would be a little tricky to do, but I
> believe that it would still be much better than having only whole
> keyframes or 3D GOFs.

Hmm, but on the other hand, wavelet compression of images isn't that 
much better than JPEG, and JPEG uses 8x8 blocks. So it seems that 
losing a bit of locality isn't that expensive.

> I hope you use floats or fixed-point for the blurred images so
> that you won't get too many neighbors with the same values.
> That algorithm reminds me of Ray Heasman's algorithm that he
> described here in March. His images were downsampled.
> For each resolution, the motion vectors were upsampled and then
> refined. For downsampled images, he could have used the lowpass
> subbands from each wavelet decomposition, that were provided for
> free!

Yeah, I use floats everywhere in my research code. I checked the 
archives for March, but found only a link to his page, which 
doesn't give that much detail though.

> Up- and down-sampling of an image will displace features, no
> matter how good the length of the filter, while gaussian blur
> will not.

That's a good point. On the other hand, if you take Shannons work 
into account, downsampling is just another way of lowpass 
filtering, just like a Gaussian blur. So using an appropriate 
filter it might just work.

> From what I have experienced from staring like a maniac at
> wavelet subbands is that longer minification filters tend to
> "chisel" out features in the image better than short filters.
> Do you think that would be good or not, for motion estimation?

I'm not really sure how well wavelets and motion compensation work 
together anyway. It seems to me that wavelets really ignore motion. 
Before the wavelet transform you have a feature at a certain point 
in your image, and after the transform its in a certain point in 
the coefficient matrix. But it hasn't been transformed in a way 
that makes it any easier to extract the information. What would be 
cool is some transform that does do that (I think that would amount 
to a kind of wavelet transform that takes into account phase 
information in a different way, but I really don't understand 
enough of the matter yet to be able to do more that guess).

Regards,

Lourens
- -- 
GPG public key: http://home.student.utwente.nl/l.e.veen/lourens.key
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.0.6 (GNU/Linux)
Comment: For info see http://www.gnupg.org

iD8DBQE9OyuHvmNyqZHWDvURAmgdAJ48OA9lu2TKjFJtYVL68u/V92WbdgCfSLbh
iBuyXcEG+mlLq7G4E6WeDbM=
=+tkx
-----END PGP SIGNATURE-----


--- >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.