How to train latent SVM model for object detection ? It crashes using Matlab 2012b, Windows 64bit OS

Question

How can I train my own detector under Windows 7 in C++ or Matlab2012b? I do not have Linux or I do not have Mac OSX.

I actually searched for it a lot before writing here. What I got is:

I would like to train my own latent SVM model and use it in C++. I checked out the OpenCV sample, cpp/latentsvm_multidetect.cpp. It works with the existing models.

Authors' website : http://www.cs.berkeley.edu/~rbg/latent/index.html . It is mentioned that the software is tested using matlab2011a under linux and mac osx. However, I run matlab2012b under Windows7 64 bit. The code crashes very often, and I tried to fix the bugs, but I cannot cope anymore.

Is there any other source code which I can use under Windows 7 64bit?

OpenCV : http://docs.opencv.org/modules/objdetect/doc/latent_svm.html . There is no code implementation for training. Only detector can be used to detect the objects with existing models.

On the other hand, somebody else asked the same question. opencv latent svm . However, The question was closed without being answered. Sad!

Also, OpenCV and Latent SVM Detector is similar to my question. I am not satisfied with the answer, I read the same in the README file in the source code in Matlab. It does not help me to overcome the bugs in the source code, if I run under Windows 7. For this question, I wanted to add a comment to ask further, I cannot add because of my reputation (I am new here).

I think, Latent SVM is fascinating and something new to detect the objects. Also, I think this should be a general issue to train our own model easily. It would be nice if anyone can guide.

Anyway, I do not understand why the training part is not implemented in OpenCV with the detector part!

Answer 1

some c++ ones:

http://www.dubout.ch/en/coding.html

http://www.uco.es/~in1majim/proyectos/libpabod/

I have had both working in windows 7 64bit, but does require a little tweaking here and there, the first less than the later.

Also as an aside, you could always try the opencv hog detector first, to see if it is sufficient for your application, you can train it with opencv svm. I found it to be quite good for simple objects that have a well defined and distinct shape.

Cheers

EDIT: Sorry the libpabod is only test time, only the the 1st link has the training code.