topic Re: Error in codomain while rotating images in Intel® Integrated Performance Primitives

Error in codomain while rotating images

mweissbach — Mon, 10 Jul 2006 19:10:14 GMT

Hi @ all

I have an grey-value-image with a codomain of [0..125] for all pixel. After rotating and interpolating it with my own implemented function, I get an image with a codomain [0..115.06]. When I use the IPP (Version 5.1) I get a range of values [-1.74..120.0] for all pixels. In my opinion, this is not possible. The same effect takes place by using the ippi_resize-functions.

I don't know, if this is a bug or a error in reasoning the matter. My idea was, that by rotating or resizing the range of the pixelvalues remains to be constant. The interpolation does IMHO only nothing or a shrinking of the range of values of the image.

Can you give me some hints whats the reason for this effect?

Thanks a lot

best regards

M.

Re: Error in codomain while rotating images

mweissbach — Wed, 29 Nov 2006 23:57:19 GMT

Hi

I use INTER_NN, INTER_LINEAR and INTER_CUBIC. Of course InterCubic and InterLinear can change the range. The problem is, that the range should get smaller on both borders by using this interpolation. The strange thing is, that the lower border gets negative values. Is there a logical explaination of this behavior?

Thanks

M.

Re: Error in codomain while rotating images

mweissbach — Fri, 01 Dec 2006 18:47:26 GMT

Hello

> Why do you think the range should be smaller?
Because you interpolate the values. Therefore you make somewhere a meanvalue or something similar ... depends on the interpolation algorithm

> It can exceed the maxboundary as well as min extrimely (see attached picture).
In my opinion it is not possible. Especially in the case of the lower border, which becomes negative.

I think, that the computation is not correct. But I'm not totally sure. My tests give me the results I espected in contrast to the ones from IPP.

M.

Re: Error in codomain while rotating images

mweissbach — Fri, 01 Dec 2006 19:39:01 GMT

To syncronize our discussion and correct my partly erronious thougts:

It is clear, that there can be an expansion of the range while interpolating bicubic and quadratic (because of the interpolation function).

But this effect also appears with the modes of no or linear interpolation.

So in this case, I think, the result is not correct ...

Thanks

M.