Clarity required on the sub-pixel accuracy

[Drakula01]

In the _get_precision function in imreg.py, it is mentioned that "we get subpixel precision now" and we take Dt = 0.25. Can there be a more elaborate explanation of how we get error margin of 0.25 pixels? I wasn't able to find anything after going through the code.

[matejak]

I think that https://imreg-dft.readthedocs.io/en/latest/cli.html#caveats describes it quite accurately - if you want subpixel resolution of the phase correlation, then you need "zoomed in" spectrum, which means that you need to expand the field of view as much as possible, while keeping the scale of the image (i.e. pixel/m ratio) the same.

Imreg_dft doesn't do any super clever magic. I believe that what you are really asking about is https://github.com/matejak/imreg_dft/blob/master/src/imreg_dft/utils.py#L343 - instead of using the plain argmax to find out where is the power spectrum strongest, the function calculates something like the center of gravity. I believe that I got better results back then, and you can try it on synthetic data to validate that hypothesis.

Now I must admit that I don't know when the _argmax_ext function is used, as the _argmax2D seems to be preferred, and I don't recall anything any more, but you should be able to verify that quite easily, as I imagine that you use the module through its bindings, and not through the command-line.