Moderator: Computer Vision

klein&#946;eta
Neuling
Beiträge: 4
Registriert: 14. Apr 2006 00:12

Hallo, what is meant by "correctly normalized"

if I normalize the filter I will end up with one that only has positive values, and thus will not compute central differences any more.

I guess simply using the sobel filter
[1 2 1;
0 0 0;
-1 -2 -1]
is not what you want here as its not exactly the values for a gaussian.

so what should we normalize the filter to? should rows sum up to 1 (or -1) or should the maximum value of a row be 1 or something else?
o.O- p

Toobee
Sonntagsinformatiker
Beiträge: 225
Registriert: 7. Apr 2011 12:58

### Re: Ex1 Task3

no, you shall use the central difference, which is
(1/2) * [1,0,-1] (see slides of previous year if they haven't upload them yet)

next you have to come up with the gaussian filter in x direction. Combine the 3x1 and 1x3 filter and you should have the needed 3x3.

qgao
Moderator
Beiträge: 33
Registriert: 7. Apr 2009 11:38

### Re: Ex1 Task3

klein&#946;eta hat geschrieben: so what should we normalize the filter to? should rows sum up to 1 (or -1) or should the maximum value of a row be 1 or something else?
sum to 1 for the Gaussian filter.

dummdidumm
Windoof-User
Beiträge: 39
Registriert: 28. Apr 2010 18:49

### Re: Ex1 Task3

Another question regarding task 3: It says "For every edge candidate pixel, consider the grid direction (out of 4 directions) that is “most orthogonal" to the edge". Does this mean that we only have to care about all horizontal and vertical edges and can discard all other edge-vectors?