Add outright support for sub-pixel precision

[matejak]

This can be achieved by doing smart things when evaluating the phase correlation.

[arve0]

Might look into scikit-image, which has implemented this: https://github.com/scikit-image/scikit-image/blob/master/skimage/feature/register_translation.py#L195

They use the reference Manuel Guizar-Sicairos, Samuel T. Thurman, and James R. Fienup, "Efficient subpixel image registration algorithms," Optics Letters 33, 156-158 (2008).

[arve0]

Just stubled upon A study of sub-pixel motion estimation using phase correlation, which uses fitting of a sinc function to the peak in the phase correlation. It claims to perform better (both resource and time wise) than up-sampling-techniques.

[matejak]

I have added linear interpolation to the log-polar stuff. It mostly improves things, but not always. Depends on #11

[matejak]

It seems that the resampling way is more robust than the interpolation.