Efficient stack reconstruction and averaging

[sk1p]

Example workflow:

• reconstruct single hologram, as the registration reference
• determine a (large enough) sample region, on which the registration is performed
• pass the registration crop into the UDF (maybe in fourier space)
• directly after reconstruction, in process_frame, determine the shift for each hologram
• the result should be both the hologram stack and the shift vectors
• run normal reconstruction (without registration) on the reference data set
• divide the not-yet-shifted hologram stack by the reference ("zipped" together)
• on the corrected hologram stack, apply shifts and generate average
• apply phase unwrapping and visualize

This should be implemented with GPU support, especially the phase correlation, as that is quite compute heavy. Should include upsampling for subpixel accuracy.

If necessary, implement different registration reference methods. For very large stacks, finding shifts relative to the beginning of the partition could work (would need to do some work in merge for global registration).

Future

Can do the averaging in the UDF, too, if we could zip together two data sets (obj + ref), basically making a (nav, 2, sig_y, sig_x) shape virtual data set. Then the UDF could directly output the averaged result.