Histogram backprojection

[Jérôme Kieffer]

On Thu, 11 Apr 2013 23:41:35 +0200
Stéfan van der Walt <stefan at sun.ac.za> wrote:

> On Apr 11, 2013 8:49 PM, "Jérôme Kieffer" <google at terre-adelie.org> wrote:
> >
> > On Wed, 10 Apr 2013 04:13:52 -0700 (PDT)
> > abid rahman <abidrahman2 at gmail.com> wrote:
> >
> > > Doesn't scikit-image has a 2d convolution function like cv2.filter2D in
> > > OpenCV which enables convolution of an image (or any numpy array) with
> > > custom kernel ?
> >
> > Convolution can be done in Fourier space, especially if the kernel is
> large ... I do that often, especially on nvidia GPU where FFT is for free.
> 
> SciPy also provides fftconvolve, for what it's worth.

I was surprised to figure out that fftpack (from numpy) can be  faster than fftw depending on the input size...
I guess it is not more than: numpy.fft.ifft2(numpy.fft.fft2(A)*numpy.fft.fft2(B).conjuguate())
 
> Looking forward to seeing those opencl codes! (We have to reopen the
> backend discussion soon)

If cuda is an option, then I have implementations based on cufft which are really efficient (but most of us could do it so trivial it is). I consider openCL is both cleaner and more promising for open source (even if less easy)


Cheers,
-- 
Jérôme Kieffer <google at terre-adelie.org>