---------- Forwarded message ----------
 From: Josna Rao <josnar@ece.ubc.ca>
 Date: Fri, May 7, 2010 at 11:56 AM
 Subject: Re: Measuring reliability
 To: Tricia Pang <triciajpang@gmail.com>
 Cc: salima.elm@hotmail.fr


 Oh neat, that's pretty close to what we're doing. We determine the
 image reliability for every pixel in the image based on edge strength,
 noise, and image curvature + texture, and then we weight the contour
 with those local measures.  I think one of Ghassan's phd students,
 Lisa Tang, is doing a project with Krishna on using our reliability
 weights for image registration, you may want to ask Krishna about that
 since I think it might be quite similar.

 In terms of noise measures, we used a pretty simple gradient measure,
 just the max(\nabla I_x, \nabla I_y) which is a basic gradient measure
 with the max term to avoid discritization issues. Incorporating the
 gradient direction is useful since you can enforce linked edges. From
 experience, the 2nd order gradient measures are more sensitive to
 noise. We 'gated' our edge cue with the noise measure so regions of
 high noise have a low reliability regardless of the edge measure. I'd
 say canny is one of the strongest edge measures, along with the
 laplacian of gaussian.

 We also measured the level of white noise in the spectral domain and
 used a curvature/corner estimator to give high curvature regions
 higher reliability. We use the scale-invariant curvature measure from
 Lindeberg: http://www.nada.kth.se/~tony/earlyvision.html. The harris
 corner detector is similar, but with Lindeberg, you don't just measure
 corners, you get a continuous measure so that slightly curved but not
 as sharp regions are given higher reliability too. We also used our
 noise measure to disregard curvature in high noise areas.

 Hope that helps!

 - Josna

 On Wed, May 5, 2010 at 5:06 PM, Tricia Pang <triciajpang@gmail.com> wrote:
 > Hi Josna,
 >
 > Hope your thesis writing is nearing an end! When you've got a sec, could you
 > give us some suggestions or feedback? I think a part of the project I'm
 > currently working on has some overlap with your work, assuming I didn't
 > totally misunderstand what your research is on. Salima (Sid's summer
 > student) is helping me out with this (she's cc'd on this email).
 >
 > A bit of background -- after drawing 2d livewire contours on slices in a 3D
 > MRI dataset, the contours are used to guide registration with an atlas
 > model. We want to calculate edge strength/noise measurements for each point
 > on the contour, and then use these to compute a "reliability/confidence
 > weight" for each contour point (eg. larger weight where there are stronger
 > edges). This then weights each point accordingly in the registration energy
 > cost function.
 >
 > To compute the weight, we were thinking of a weighted sum of measurements
 > given by gradient and contrast (Haralick feature) information... but then
 > there are the other usual edge detection algorithms (canny, sobel, harris
 > corner, laplacian of gaussian, difference of gaussians, determinant of
 > hessian, etc) that we haven't really considered yet.
 >
 > From your experience, what methods have worked, and do you have suggestions
 > for any other algorithms that we should look into?
 >
 > Thank you!!!
 > Tricia
 >