Comments on: OpenCV Haartraining: Detect objects using Haar-like features

By: shri

shri — Fri, 03 Sep 2010 07:48:13 +0000

thanks for the reply,i got the answer from the function icvCreateCARTStageClassifier which is in cvhaartraining.cpp, ( left val and right val and threshold are adaboost algo dependent)

about the haar feature value: sum of black rectangle value -sum of white rectangle ,

depending on the values the sum will be -ve or +ve.

thanks once again for the reply

By: Andrey Soldatov

Andrey Soldatov — Thu, 02 Sep 2010 01:27:34 +0000

Hi shri, I’m not sure regarding the question. Please, clarify ‘can it be -ve or +ve (both)’. Value of Haar feture can be positive or negative because weight of a rectangle can be positive or negative. After comparison of feature with threshold the algorithm chooses one of positive values: ‘left_val’ or ‘right val’.

By: Andrey Soldatov

Andrey Soldatov — Thu, 02 Sep 2010 01:14:41 +0000

I’m not sure I understood you. If you just want to use haartraining, all these rules are inside OpenCV algorithm, so you don’t actually need to know anything of them. Please, confirm if you need some more explanation about internals of the algorithm.

By: Tobi

Tobi — Thu, 19 Aug 2010 10:06:02 +0000

Hi Chenwei,

I also had the same problem! I configured everything correctly, but still I had empty output. How I solved my issues:
I found another compilation of my openCV and used its opencv_performance and all of a sudden it worked out. Maybe you try that too! (I know it is no solution, but it is a point to start)!!

greets, Tobi

By: shri

shri — Tue, 03 Aug 2010 11:52:14 +0000

some one please tell how to get the threshold and left val and right val as explained by anrey, as am new to haar training.

By: shri

shri — Sat, 17 Jul 2010 10:28:35 +0000

hi andrey, please give some more explanation regarding weak classifier and its error calculation ,

ie, the selection of weak classifier, by AdaBoost

then the value of Harr feature can it be -ve or +ve (both)?

thank you in advance

By: Andrey Soldatov

Andrey Soldatov — Fri, 09 Jul 2010 05:58:28 +0000

In general, you need some knowledge about the Viola-Jones object detection algorithm and Haar features. You can start from wiki pages or from http://opencv.willowgarage.com/documentation/c/object_detection.html. There is short description of Haar features. Or read http://research.microsoft.com/en-us/um/people/viola/Pubs/Detect/violaJones_CVPR2001.pdf. But simple description of the parameters is as follows.

‘feature’ – a rule to get one number from multiple points of image. Object detection algorithm gets several (or several hundreds) such numbers from image and then analyzes them to return answer if it can detect the object.

Haar features are calculated based on two rectangles on image (To be more precise, on window of some size got from Viola-Jones algorithm. This size is encoded in xml in ‘size’ field.) as sum of intensity of pixels on each rectangle. Then results for two rectangles are summed with weights (sum = sum1 * weight1 + sum2 * weight2). 5 numbers for each rectangle mean: (x, y, width, height, weight). Pay attention that the coordinates are related to window and not to whole image.

Then the algorithm compares sum with ‘threshold’ and decides if feature is ‘pro’ or ‘con’ to say that an object detected. As a result of the comparison it selects ‘left_val’ or ‘right val’. negative value is ‘con’, positive is ‘pro’.

On each stage the algorithm sums values ‘pro’ and ‘con’ for several features and compares this sum with stage_threshold. If sum is lager than ‘stage_threshold’, window is considered as passed the stage.

If some window passed all stages it passed whole classifier.

By: Jsay

Jsay — Tue, 22 Jun 2010 12:51:02 +0000

Tags like “feature”, “rects”, “threshold”, “light_val”, “right_val”…

By: Jsay

Jsay — Tue, 22 Jun 2010 12:49:45 +0000

Hello,

I’m just wondering that how to interpret the resulting xml-file? There are tags like , , , , , ..

Could you explain me those and the values included?

By: Andrey Soldatov

Andrey Soldatov — Wed, 16 Jun 2010 01:06:12 +0000

Hi, Atul. I believe, you use incorrect (too small) size of sample. You should replace “20 20″ with the size used during training.

But another question, for which purpose do you use haarconv? OpenCV creates xml file in the end of training and there is no need to make it manually. Besides, OpenCV package contains convert_cascade.exe if you need to convert temporary results of training to xml to perform some tests while cascade is not ready yet. In both cases of haarconv and convert_cascade you must specify correct size of sample.