Hello hello,
I started this absolute car crash of a PR because we’ve seen significant speed improvement for our uint16 images when calling phase_cross_correlation (real space, upsample=20) and casting to float32. Now, I don’t fully understand what is going on, but our tests showed no significant loss in accuracy.
I am working on writing a sane benchmark as opposed to my silly timeit test. But meanwhile, I thought I’d start a conversation here, in case anyone has some thoughts about why this is or is not a good idea.