The correlation is computed at pixel level, similar to step COARSECORR. These computations are described in that chapter. (That step still has to be performed because the FINE step requires accurate initial estimates of offsets. The AccL and AccP cards define the size of the searchwindow (2*AccL x 2*AccP) around the initial offsets to interpolate a maximum.)

The oversampling is done as follows:

- Transformation to spectral domain of searchwindow with correlation values at pixel level.
- Padd with zeros, half the last term.
- Inverse transform.
- Find maximum in space domain, this corresponds to estimated offsetvector.

Note that this way of computing is exactly the same if you first
interpolate the signal and compute all correlations and find the
maximum, or that you first compute at pixel level and interpolate the
correlation values.

Leijen 2009-04-14