image processing - Harris Laplacian Detector and auto scale selection -

i want basic understanding of harris laplacian detector. know harris detector cannot figure out idea behind harris laplacian detector.

let's assume have 2 images showing pretty same thing 1 image scaled. want find interest points can check them later correspondance , compute homography matrix later. need scale invariante interest point detector.

what think harris-laplacian works (very abstract; correct me if i'm wrong):

• apply harris detector both images
• apply function each interest points returned harris detector
• keep points maximize laplacian
• scale images , apply harris detector again both images
• ...repeat other steps...

my questions:

1. what kind of points maximizes laplacian?
2. the function apply interest points, like? apply interest points or patch around points?
3. i've read maxima of function used. need for, tell us?
4. it seems method kind of bruteforce (scale images, apply harris, check points keep, scale images again, apply harris, ...). isn't inefficient?
5. the points want keep have maximize laplacian. 'the laplacian'? laplacian operator?
6. if point maximizes laplacian have couple of points each scale. how final interest points - scale them?
7. when algorithm terminate?
8. why use laplacian?

pretty many questions :/

i think best if read mikolajczyk et al paper harris-laplace , maybe parts of thesis; try answer as questions..

first, why asking points maximize laplacian ? 3rd derivative , 4th derivative of laplacian give information extremas of laplacian, not case in here. point maximizes on laplacian scales used in harris laplace detector.

second, "apply function each interest points returned harris detector", far know, no don't apply function. so;

1- not maximize laplacian, maximize on laplacian scales found in algorithm f[laplacian(scale = x-1)] < f[laplacian(scale = x)] , f[laplacian(scale = x)] > laplacian (scale = x+1),this point accepted maxima in laplacian scale in manner. reject points values lower threshold defined. laplacian value defined scale scale , point x; f(scale = i)=sigma(i)^2* |(laplacianxx(x,i) * laplacianyy(x,i))|

2- function ? please comment if missing point in here.

3- since don't know function, don't know maxima.

4- gaussian , derivative linear function. therefore, taking derivative of gaussian , laplacian, simple convolution of image filter provides log of image. however, there other scale invariant point detectors many of can use if think perform faster case.

5- yes laplacian think answered maximum confusion already.

6- them if don't know scaling.otherwise, ,f ask choosing scales , patch sizes image matching, scale matching. laplacian measure between scales, similar values. @ first slides of presentation http://campar.in.tum.de/twiki/pub/chair/teachingws09matdcv/feature_detectors2.pdf. in addition may want read in thesis part titled, 'scale covariant points'. find link @ bottom.

7- decide limits of scale space. after program finishes calculating scales defined(and of course points detected), program should finish. not talking 2 images , corresponding problems yet, should concentrate on later.

8- in paper, found laplacian scale more effective 3d-scale space. stackoverflow question harris-laplacian-detector: corner- , blob-detector?. find thesis in explains harris-laplace detector clearly.

i hope helps, luck :)