🚀 go-pugleaf

I'm currently engaged in a project doing vergence stereo.
For disparity estimation I'm using a cepstrum filter, as
proposed by Yeshurun & Schwartz, Olson & Coombs etc. (se below).
I'm working on 32x32 images, downsampled from 256x256.

When doing the cepstrum transformation, only a small bit
of the transform is really interesting, namely the bit
containing the disparity peak. So, my plan was to use
zoom-fft in the cepstrum transform, so only the interesting
bit is actually calculated, and with greater resolution too.

For people not acquainted with zoom-fft, it's actually just
multiplication of a linear phase factor in the spatial
domain, to perform a shift in the frequency domain. If this
is done before downsampling (e.g. on the 256x256 images),
the resulting 32x32 fft will be a 'close-up' of the
frequency range centered by the frequency shift.

So far, all I've had is failure. My Matlab calculations
do the frequency shift ok, but if I just do a real cepstrum,
the interesting echo peaks just vanish(!).
Trying to do a complex cepstrum gives me big trouble doing
phase unwrapping of the log function. This is no problem
in 1D, but in 2D? 


Having thus described my problem, I call upon the experts
of the net:
  Are my efforts futile, or is 2D phase unwrapping possible?
And if it is, will the zoom-cepstrum transform contain the
desired disparity peak, or does the zooming eliminate the
correlation information?


Holding my breath...
--Anders.


Litterature:

	"Cepstral Filtering on a Columnar Image Architecture:
	 A Fast Algorithm for Binocular Stereo Segmentation"
	Yehezekel Yeshurun and Eric L. Schwartz
	IEEE Trans. on patt. anal. and mach. int. Vol.II,no.7,july 1991

	"Real-Time Vergence Control for Binocular Robots"
	Thomas J. Olson and David J. Coombs.
	Intl. Journ. of Comp. Vis. 7:1, pp.67-89 1991

--

Anders Pedersen.   Dept. for Image Analysis, Aalborg University, Denmark
     anders@tuborg.nork.auc.dk  ... or ...  i8anders@aud2.aud.auc.dk

                      Disclaimer:    Intel inside.