Fabien R's profile - activity

A possible solution is to initialise the variables before calling cv::fisheye::calibrate():

rvecs.resize(imagePoints.size(), cv::Mat(1, 1, CV_64FC3));
tvecs.resize(imagePoints.size(), cv::Mat(1, 1, CV_64FC3));

A sample of the undistortion after calibration:

I'm trying to calibrate a fisheye camera using opencv3. But it crashes in cv::fisheye::calibrate() in fisheye.cpp on this line:

  rvecs.getMat(i)=omc[i];

It seems that rvecs is empty, so getMat() fails. I would like to resize rvecs to contain the number of elements of omc. I tried

 if (rvecs.empty())
{
    std::vector<cv::Mat>* v = (std::vector<cv::Mat>*)&rvecs;
    v->resize(omc.size());
}

But it crashes because OpenCV dereferences things when closing the bracket.

Any hints ?

@crazyfffan: That's the main issue with the OpenCV implementation. There's no function to monitor the learning phase. It would be helpful to plot the curves related to bias/variance problems.

Thanks berak. For the moment, I think I will stick to what everyone does, i.e., removing a part of the training sample set to feed the testing process.

Sorry for the short post. I edited the question.

I'm trying to generate test samples for MLP (OpenCV 2.3).

I found the function cvCreateTestSet in ml.hpp but I did not see explanations about how to use it.

Any hints ?

Edit: I would like to adapt this excellent wiki by replacing kMeans by MLP(back_prop). I was wondering if cvCreateTestSet could be used somewhere in the process.