kabamaru's profile - activity

The linker needs to know where your header files are and it tries to find them by looking to some places that he already knows. You need to tell him to look where you headers are. For this you have to set two environment variables C_INCLUDE_PATH (for C header files) or CPLUS_INCLUDE_PATH (for C++ header files). You can do this with the EXPORT command. A nice practice is to put this command in your .bashrc file so you don't have to do the same every time. The directory for the opencv files depend on where you've installed it but generally are /usr/local/include and /user/local/include/opencv/. Probably you will also need to set the LD_LIBRARY_PATH.

Another option is to use pkg-config with the compiler.

gcc example.c `pkg-config --libs --cflags opencv`

Your code seems wrong: deviationframe= deviationframe.mul(deviationframe);

Should be something like:

bufferFrame =(frame32f - resultframe);
deviationframe += bufferFrame * bufferFrame;

It's probably because the multiplication returns a mat and not a float. Tried accessing it with at<float>(0,0)?

In opencv 2.4 and beyond there are two implemented alternatives:

A part based model: http://docs.opencv.org/modules/objdetect/doc/latent_svm.html#latentsvmdetector

HoG, which can be found under the objdetect module. It's not very well documented though. It also has a GPU implementation: http://docs.opencv.org/modules/gpu/doc/object_detection.html

Try

hogGPU.computeConfidenceMultiScale(imgGpu, found, 0, Size(8,8), Size(0,0), confidenceScores,2);

:)

It's not that I don't want to answer but it's really annoying because instead of answering you just "polute" this thread. In any case if you want an answer please post a question.

You must be kidding me right? Your answer is a different question? What kind of breaking the etiquette is that?

I use HoG to detect pedestrians in a hybrid system and I need to use the posterior probability from the classifier.

In libsvm it is possible to train the classifier to return scaled probabilities (uses Platt's method).

In openCV 2.4.3 in the gpu module under the HOGDescriptor class, there were added two functions (computeConfidence and computeConfidenceMultiscale) that return confidence of classification instead of binary class.

Since opencv uses a version of libsvm does anyone know if these confidences are are already scaled? I couldn't find any relevant documentation.