[Numpy-discussion] numarray bug in gaussian_filter1d?

[Peter Verveer]

I think you are correct: The result of a gaussian filter of order one  
should be a derivative operator. Thus the response to an array  
created with arange() should be close to one (barring edge effects).  
Currently we have:

 >>> from numarray import *
 >>> from numarray.nd_image import gaussian_filter
 >>> a = arange(10, type = Float64)
 >>> gaussian_filter(a, 1.0, order = 1)
array([-0.36378461, -0.84938238, -0.98502645, -0.99939268, -0.999928  ,
        -0.999928  , -0.99939268, -0.98502645, -0.84938238,  
-0.36378461])

So the sign is wrong, that can be fixed by mirroring the gaussian  
kernels. I have done so in CVS. The same holds true for the Sobel and  
Prewitt filters, they were also defined 'incorrectly' according to  
this criterion. I also changed those. That may be a bit more  
controversial since a quick look on the web seemed to indicate that  
often it is defined the other around. If anybody thinks my changes  
are no good, please let me know.

Cheers, Peter

On 22 Sep, 2005, at 17:41, Alexandre Guimond wrote:

> Hi.
>
> I think I found a bug in gaussian_filter1d.
>
> roughly the function creates a 1d kernel and then calls  
> correlate1d. The problem I see is that the kernel should be  
> mirrored prior to calling correlate1d since we want to convolve,  
> not correlate.
>
> as a result:
>
> >>> import numarray.nd_image
> >>> numarray.nd_image.gaussian_filter1d( ( 0.0, 1.0, 0.0 ), 1,  
> order = 1, axis = 0, mode = 'constant' )
> array([-0.24197145,  0.        ,  0.24197145])
> >>>
>
> when it should be [ 0.24197145,  0.        ,  -0.24197145]) (notice  
> the change in the sign of coefficients)
>
> Or did I get that wrong?
>
> alex.