I was rereading these notes and got a little confused about what G is on this slide, can I get some clarification?
mpotoole
$G$ would be some kernel in this case. For example, for step 1, $G^x_\sigma$ might represent the derivative of a Gaussian in the x- direction.
adigrao2012
Where did the subtraction of the mean go?
mpotoole
@adigrao2012 Ahh really good question! Where did the mean go? The corner detection procedure shown here seems slightly different from the one shown on this slide.
I suppose the answer is that there are two slightly different ways to interpret corner detection: (i) one that involves performing PCA, which technically requires subtracting the mean, and (ii) one that involves minimizing this error function, which does not require subtracting the mean (see Quiz 2). The latter describes the procedure detailed in the original Harris corner detector paper.
So what are the implications of not subtracting the mean? The procedure here may have some trouble with gradients, because the error function would give a non-zero response in such regions. I suspect that subtracting the mean is a good idea in practice---though I admit that I'm not entirely sure.
I was rereading these notes and got a little confused about what G is on this slide, can I get some clarification?
$G$ would be some kernel in this case. For example, for step 1, $G^x_\sigma$ might represent the derivative of a Gaussian in the x- direction.
Where did the subtraction of the mean go?
@adigrao2012 Ahh really good question! Where did the mean go? The corner detection procedure shown here seems slightly different from the one shown on this slide.
I suppose the answer is that there are two slightly different ways to interpret corner detection: (i) one that involves performing PCA, which technically requires subtracting the mean, and (ii) one that involves minimizing this error function, which does not require subtracting the mean (see Quiz 2). The latter describes the procedure detailed in the original Harris corner detector paper.
So what are the implications of not subtracting the mean? The procedure here may have some trouble with gradients, because the error function would give a non-zero response in such regions. I suspect that subtracting the mean is a good idea in practice---though I admit that I'm not entirely sure.