Hi, I see you haven't benchmarked against the skimage.transform.* functions, those should be a lot faster compared to `ndimage.map_coordinates`. If your implementation provides a different kind of interpolation, I suggest to integrate them into the current functions and add a new parameter `mode="fourier"`. I'm not sure if there is an equivalent for your functions in all cases (see `transform/_warps.py`). Thoughts? Am 23.12.2012 um 08:30 schrieb Adam <keflavich@gmail.com>:

Benchmark tests created: https://github.com/keflavich/image_registration/blob/master/examples/benchma... https://github.com/keflavich/image_registration/blob/master/examples/benchma...

They ought to be more extensive, but give the general results: -fourier-based shifting is a little faster (maybe 50%), but follows the same power-law -fourier-based upsampling is marginally slower, but my test results so far are questionable because of machine variability

Of course, with fftw3, one can increase the number of processes used for the fft, which may speed things up.

On Sat, Dec 22, 2012 at 8:31 PM, Stéfan van der Walt <stefan@sun.ac.za> wrote:

...

On Sat, Dec 22, 2012 at 5:24 PM, Adam Ginsburg <keflavich@gmail.com> wrote:

...

Reasonable idea. I think they are faster for at least some cases, but they also behave a little differently than other interpolation techniques. So put them all in a "fourier_interpolation.py" file, then put them somewhere in transforms?

Yes, I think some benchmarks would be interesting. Otherwise, we may want to add sinc interpolation to ndimage and see if that yields the same results.

Thanks! Stéfan

--

-- Adam

--