[Numpy-discussion] nd_image.affine_transform edge effects

Zachary Pincus zpincus@stanford....
Fri Mar 23 11:20:09 CDT 2007


Hello all,

On Mar 23, 2007, at 3:04 AM, Stefan van der Walt wrote:

> On Thu, Mar 22, 2007 at 11:20:37PM -0700, Zachary Pincus wrote:
>> The actual transform operators then use these coefficients to
>> (properly) compute pixel values at different locations. I just
>> assumed that the "pre-filtering" was broken (even on natural images
>> with smooth variations) because images pre-filtered with the spline
>> filter didn't look like traditional pre-filtered images ought.
>>
>> IMO, ticket 213 should be closed as PEBCAK (I'll do that forthwith);
>> further I'm sorry to have caused Peter to be further bothered about
>> this non-issue.
>
> The ticket is more concerned with the usability of ndimage.
> Currently, the principle of least surprise is violated.  If you use
> the default arguments of ndimage.rotate (except for axes, which should
> still be fixed to be (1,0) by default) and rotate Lena by 30 degrees,
> you get something with green splotches on (attached to the ticket).
>
> IMO, we should either change the default parameters or update the
> documentation before closing the ticket.

Hmm, this is worrisome. There really shouldn't be ringing on  
continuous-tone images like Lena  -- right? (And at no step in an  
image like that should gaussian filtering be necessary if you're  
doing spline interpolation -- also right?)

I assumed the problem was a non-issue after some testing with a few  
natural images didn't introduce any artifacts like those, but now  
with the Lena example I'm not sure. This is frustrating, to be sure,  
mostly because of my limited knowledge on the subject -- just enough  
to be dangerous.

Zach





More information about the Numpy-discussion mailing list