1
$\begingroup$

Harris and Stephens writes about the interest window of Moravec: "The response is noisy because the window is binary and rectangular", and suggests applying a Gaussian window.

My Question: Why is the response not noisy after applying a Gaussian window? Is it because a Gaussian filter removes noise from a picture, or because a circular window is better for sampling somehow?

Thank you.

$\endgroup$
  • $\begingroup$ Your question makes little sense to me. 1) What is the Moravec window? 2) What is a "binary rectangular" window? 3) If the Euclidean distance doesn't vary from the centre to the edge, then the centre IS the edge (i.e. there is no distance). 4) Harris solves what by applying a Gaussian window? 5) Nothing before your question shows that someone is saying that a "circular window is less responsive to noise than a rectangular window" where did that come from? 6) "the Euclidean distance from the center to the edge of the window should be the same" as what? I have not had my coffee yet! :-) $\endgroup$ – Peter K. Nov 25 '15 at 12:42
  • 1
    $\begingroup$ I have tried to clear up the question and not go into details which i don't understand. Thank you for pointing it out. $\endgroup$ – user89423 Nov 25 '15 at 12:58
  • $\begingroup$ Thanks for clearing that up! That makes more sense (and I've now had my coffee). $\endgroup$ – Peter K. Nov 25 '15 at 13:09
  • 1
    $\begingroup$ Wouldn't it be called "Moravec interest operator", or "Moravec corner detector" from H. P. Moravec, Visual Mapping by a Robot Rover, 1979? $\endgroup$ – Laurent Duval Nov 25 '15 at 13:14
0
$\begingroup$

I've added a link to the paper you're quoting in the question.

The issue is that both the Moravec corner detector, and the change to it suggested first in that paper:

enter image description here

use the derivative of the image. ANY difference operation will tend to enhance higher frequency noise in the image. The idea with the Gaussian smoother is to reduce this effect. The two operations (Gaussian smoothing and difference operation) are often combined into a single operation: the difference of Gaussians.

As the paper suggests, the anisotropy of the Moravec detector is why the differential is used, rather than the up, down, and diagonal differences in the Moravec algorithm. The smoothing is introduced to counter the noise issue.

$\endgroup$
  • $\begingroup$ Thank you Peter that helped a lot. I have a follow up question regarding anisotropy. Isn't the response of Harris still anisotropic as the derivatives considered only are vertical and horizontal? $\endgroup$ – user89423 Nov 25 '15 at 13:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.