Thread Subject:

multi peak surface 2D Gaussian fitting

I have a 2D data, xy plane is my pixels and z axis is the values corresponding to each pixel(like an image).
I want to model my z data with Gaussians, how can I cluster my data and fit multi peak Gaussians?
Is there such a package of Matlab or any other way, code etc.?

Thanks a lot

On Aug 9, 4:03 pm, "pasino pasin" <ozlempa...@hotmail.com> wrote:
> I have a 2D data, xy plane is my pixels and z axis is the values corresponding to each pixel(like an image).
> I want to model my z data with Gaussians, how can I cluster my data and fit multi peak Gaussians?
> Is there such a package of Matlab or any other way, code etc.?
>
> Thanks a lot

-----------------------------------------------------------
Are your Gaussians well separated, or are they mixed? If they're well
separated, you can use the CLEAN algorithm used in astronomy to find
stars:
http://www.cv.nrao.edu/~abridle/deconvol/node7.html
I've talked about it before - explained the algorithm. Search the
newsgroup for my earlier messages.

ImageAnalyst <imageanalyst@mailinator.com> wrote in message <e63ae1b1-a6b6-4f88-8ac7-85a9a19dbb98@w30g2000yqw.googlegroups.com>...
> On Aug 9, 4:03 pm, "pasino pasin" <ozlempa...@hotmail.com> wrote:
> > I have a 2D data, xy plane is my pixels and z axis is the values corresponding to each pixel(like an image).
> > I want to model my z data with Gaussians, how can I cluster my data and fit multi peak Gaussians?
> > Is there such a package of Matlab or any other way, code etc.?
> >
> > Thanks a lot

Actually, I had the same/similar problem and ended using the Isophote curvature, i.e. analysing the gradient direction and curvature at each pixel to detect the peaks of Gaussian as well as to segment them; as described in "Multimodal Saliency-based Attention for Object-based Scene Analysis" (by B. Schauerte, B. Kühn, K. Kroschel, R. Stiefelhagen). One advantage of this approach is that it is fast (with constant runtime) and can deal with some overlap of the Gaussians as well as with some noise. However, if you implement it, you really need to take care about good filtering when calculating the necessary image derivatives.

Hi Boris,

What "good filtering" did you use when calculating the necessary image derivatives?

Thank you!

"Boris Schauerte" wrote in message <j7scta$bbm$1@newscl01ah.mathworks.com>...
> ImageAnalyst <imageanalyst@mailinator.com> wrote in message <e63ae1b1-a6b6-4f88-8ac7-85a9a19dbb98@w30g2000yqw.googlegroups.com>...
> > On Aug 9, 4:03 pm, "pasino pasin" <ozlempa...@hotmail.com> wrote:
> > > I have a 2D data, xy plane is my pixels and z axis is the values corresponding to each pixel(like an image).
> > > I want to model my z data with Gaussians, how can I cluster my data and fit multi peak Gaussians?
> > > Is there such a package of Matlab or any other way, code etc.?
> > >
> > > Thanks a lot
>
> Actually, I had the same/similar problem and ended using the Isophote curvature, i.e. analysing the gradient direction and curvature at each pixel to detect the peaks of Gaussian as well as to segment them; as described in "Multimodal Saliency-based Attention for Object-based Scene Analysis" (by B. Schauerte, B. Kühn, K. Kroschel, R. Stiefelhagen). One advantage of this approach is that it is fast (with constant runtime) and can deal with some overlap of the Gaussians as well as with some noise. However, if you implement it, you really need to take care about good filtering when calculating the necessary image derivatives.