HOGDescriptor

[pefi9]

Hi,
I can't get HOG descriptors, respectively it's all zeros.

 opencv_objdetect.HOGDescriptor hogDesc = new opencv_objdetect.HOGDescriptor(
            new Size(64, 128)
            , new Size(16, 16)
            , new Size(8, 8)
            , new Size(8, 8)
            , 9);

 float[] desc = new float[(int)hogDesc.getDescriptorSize()];
  Mat temp = cvarrToMat(img);
  hogDesc.compute(temp, desc);

[saudet]

Have you tried to do the same thing as shown in this OpenCV sample?
https://github.com/Itseez/opencv/blob/3.0.0/samples/tapi/hog.cpp

[pefi9]

My understanding of this sample is that it`s for detecting people with pre-trained SVM classificator:

hog.setSVMDetector( HOGDescriptor::getDaimlerPeopleDetector() );

I`m trying to create simple OCR where HOG descriptor would provide me gradients of the different parts of an image and svm would be trained based on these descriptors.
More like the 2nd answer here: http://answers.opencv.org/question/10374/how-to-training-hog-and-use-my-hogdescriptor/

[saudet]

Ok, so would you have some code in C++ that is known to be working? And if you cannot find anything, please ask the authors of HOGDescriptor for usage guidelines. It's not like I'm supposed to know how to use that undocumented API.

[pefi9]

Sure, thanks.

[saudet]

I've tried this here and it returns something else than zeros:

    Mat image = imread("/home/saudet/projects/bytedeco/javacv/samples/pic1.png");
    HOGDescriptor hog = new HOGDescriptor();
    FloatPointer descriptors = new FloatPointer();
    hog.compute(image, descriptors);

But descriptors.limit() is much larger than hog.getDescriptorSize(), so I guess you're only getting the beginning of the data. You need to specify a larger array if you want to use float[] instead of FloatPointer.

[pefi9]

I see, thanks a lot.

[pefi9]

I tried it with "FloatPointer" instead of float[]. The size is just the same however FloatPointer is populated with number after calling hogDescriptor.comput while float[] is empty (all zeros).

Anyway, FloatPointer works - thanks for the hind.

[saudet]

@pefi9 Yes, it seems there's a problem there... I'll have to look into that.

[pefi9]

Hi @saudet,
how can I load trained svm model? I saved the model by using svm.save(path), but I can`t figure out how to load it (something like svm.load(path)?).

[saudet]

What do they do in the C++ samples for that?

[pefi9]

void CvStatModel::load(const char* filename, const char* name=0 )

Taken from:
http://docs.opencv.org/2.4/modules/ml/doc/statistical_models.html

Example:

CvSVM *SVM = new CvSVM;
SVM->load("SVM.xml");
SVM->predict(DataRow);

http://answers.opencv.org/question/13418/machine-learning-saveload-problem/

[saudet]

That's for OpenCV 2.4. Any examples for OpenCV 3.0?

[pefi9]

http://docs.opencv.org/3.0-beta/modules/ml/doc/statistical_models.html#statmodel-load

[saudet]

I see, they've moved the load() methods to function templates. I've added something to the commit above. Let me know if this satisfactory! I've also fixed the problem with the HOGDescriptor.compute() and int[] in the other commit. Thanks for reporting!

[taobupt]

there is still no function load() in SVM, so what should I do if I want to load a classifier file from a path. please give a hand

[saudet]

@taobupt It's all here: https://github.com/bytedeco/javacpp-presets/blob/master/opencv/src/main/java/org/bytedeco/javacpp/helper/opencv_ml.java#L48

[taobupt]

there is no opencv_ml.java in *javacv-1.1-bin.zip (140 MB), *which means
that I have to compile the source code by myself? you can add it to the
javacv-1.1-bin.zip if you have some free time. Thanks for help.

[saudet]

Please follow the build instructions here, if you have some free time:
https://github.com/bytedeco/javacpp-presets/#build-instructions

[saudet]

This has all been fixed and included in the just released version 1.2. Thanks for reporting all this guys!