Marina, could you please review this request. More details are available in the issue details on code.opencv.org.

@NCBee , why we can't use groupRectangles with signature that support weighting (objdetect.hpp, line 331)

@cuda-geek I spent a couple of hours reading through the (undocumented) groupRectangles code and at first it seemed very specific to the CascadeClassifier class. However after experimenting with it tonight, I found a "work-around" that if you pass an initialized dummy vector to groupRectangles, it gets the job done:

    // inside hog.cpp's void HOGDescriptor::detectMultiScale(...) const method definition
    // ...
    if ( useMeanshiftGrouping )
    {
        groupRectangles_meanshift(foundLocations, foundWeights, foundScales, finalThreshold, winSize);
    }
    else
    {
        std::vector<int> weights;
        weights.resize(foundLocations.size(), INT_MAX);
        groupRectangles(foundLocations, (int)finalThreshold, 0.2, &weights, &foundWeights);
    }

This is a hack and we are not using the function the way it was intended to be used (possibly by CascadeClassifier). However if you think this hack does the job, feel free to close this pull request and either you or me can add the extra two lines inside hog.cpp to fix the problem.

Note: IMO in general the location of groupRectangles() functions inside cascadedetect.cpp is problematic both logically and for future development, if groupRectangles() is going to be used by many algorithms outside of cascadetect.cpp. My groupROI was an attempt to bring sanity to hog.cpp by making it independent of cascadedetect.cpp.

By the way groupRectangles with signature declared on line 331 of objectdetect.hpp does not work simply because weight is a vector of int but SVM distances are vector of double. I used the signature on line 333, with a dummy variable for weight.

I simplified the solution with a new groupRectangles overloaded function defined in cascadedetect.cpp (just like other groupRectangles() functions).

@NCBee Looks like the 'Objdetect_HOGDetector.regression' test fails now, could you fix?
(http://build.opencv.org/builders/precommit_windows/builds/3608/steps/opencv_test_objdetect/logs/stdio)

@apavlenko I noticed earlier. The detector seems to now pick up more ROIs over the threshold. Given how simple the fix is, I think it's the original regression test that needs to change. It's possible that the test was based on incorrect weight information for each detected ROI. To be more specific, if we focus on one image out of the few in regression analysis, it is possible that the test was built using grouped ROIs of vector<Rect>::size() == 2 for that image, but probably used the first 2 weights stored in the vector<double>::size() == 16, which is the un-grouped weight vector and is obviously not holding the correct weights.

However, since this is a regression test we need to be very careful when making changes. I have downloaded opencv_extra and will run the tests locally to see why we get a regression error. Given my schedule it may take a week or two.

After going through the test cases and following the logic in groupRectangles( vector<Rect>& rectList, int groupThreshold, double eps, vector<int>* weights, vector<double>* levelWeights ) function in cascadedetect.cpp, I don't think the short method to group weights with ROIs using the current methods works. Therefore I uploaded a slightly modified version of my previous solution where a public method called groupRectangles was added to HOGDescriptor class to take care of both ROIs and weights.

Marina, could you review the updated version?

Marina,

The pull request is very old. Could you please review the patch?

👍 , It can be.