Ask Your Question
1

HOG detectmultiscale weight scale explanation

asked 2016-05-27 12:16:19 -0600

tegna123 gravatar image

Hello all,

I'm currently working with the HOG detector in OpenCV and it returns fairly good results. I only have one question, the detectmultiscale function sets the weights for each detection ROI. But what is actually the scale of good to bad? Maybe this is a really dumb question but I can't find it. I thought close to 1.0 is good but I'm getting weights of 2.0. So what is the actual scale because I thought it is from 0.0 to 1.0

edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted
5

answered 2016-05-28 05:41:32 -0600

LorenaGdL gravatar image

Quick answer: the larger, the better (the more confident)

A little bit more explanation:

The weight returned by the method for each ROI is the distance from the sample to the SVM separating hyperplane (in its corresponding kernel space). Therefore, a larger distance indicates a sample classifier with a larger confidence, as it is more far away for the other class samples. There is no upper limit for such distance, so you can't normalize and get direct probabilities from the distances. However, there are some methods to perform such convertion (e.g. the LIBSVM library implements some probability estimates, refer to its manual for details and related papers).

Also, every SVM is normalized during training, so that the distance from the separating hyperplane to each support vector sample is 1.0. Therefore, every sample with weight 0.0-1.0 lies in th SVM margin. If you want to know more about SVM in detail, refer to

Christopher J. C. Burges. 1998. A Tutorial on Support Vector Machines for Pattern Recognition. Data Min. Knowl. Discov. 2, 2 (June 1998), 121-167

edit flag offensive delete link more

Comments

So If I get this right a higher probability is further away from the SVM hyperplane, which makes it more confident. The detectmultiscale function has a threshold and maxThreshold parameter, my understanding of those parameters is that they are the minimal and maximum distance of the separating SVM hyperplane. So if I set the minimal threshold at 1, it should only return probabilities that have a distance to the hyperplane of at least 1, right?

tegna123 gravatar imagetegna123 ( 2016-05-28 10:10:50 -0600 )edit

Yes, the hitThreshold param sets the minimum distance to the separating hyperplane. I don't know which param are to referring to as the maxThreshold...

LorenaGdL gravatar imageLorenaGdL ( 2016-05-28 12:27:09 -0600 )edit

Thank you very much for the explanation! The actual name of the parameter is finalThreshold. I'm using the Java interface so maybe there is a difference in the C++ version.

tegna123 gravatar imagetegna123 ( 2016-05-29 06:52:43 -0600 )edit

The finalThreshold param refers to something quite different in fact (yep, the name might be confusing). It controls the non-maxima supression algorithm that groups similar ROIs. It is the groupThreshold param of the groupRectangles() function.

LorenaGdL gravatar imageLorenaGdL ( 2016-05-29 07:24:36 -0600 )edit

That is indeed confusing. Thank you for the explanation! I wished that OpenCV docs were a bit more user friendly, it expects you to know every detail of the algorithm.

tegna123 gravatar imagetegna123 ( 2016-05-29 11:51:23 -0600 )edit

Question Tools

1 follower

Stats

Asked: 2016-05-27 12:16:19 -0600

Seen: 2,319 times

Last updated: May 28 '16