Lucas Walter's profile - activity

I ended up following this: http://stackoverflow.com/questions/28..., though I only needed to skip the first line, the rest of my yamls are not opencv data types, just strings and numbers and dicts and lists.

It's more than just an array, it is a bunch of key value pairs. cv2.cv.Load on my yamls produce The node does not represent a user object (unknown type?) errors.

Setting <Write_extrinsicParameters> to 0 avoids the problem and allow the calibration to finish.

I also discovered I needed to bootstrap the calibration using the intrinsic guess in order to get that reprojection error down to something reasonable.

I'm trying out fisheye calibration with samples/cpp/tutorial_code/calib3d/camera_calibration/camera_calibration.cpp. I've adapted the example VID5.xml file, then calibrated without fisheye, and get a 1.36 pixel reprojection error. Next I added <Calibrate_UseFisheyeModel>1</Calibrate_UseFisheyeModel>, and get an exception:

Re-projection error reported by calibrateCamera: 282.075
Calibration succeeded. avg re projection error = 282.075
OpenCV Error: Assertion failed (rvecs[i].rows == 3 && rvecs[i].cols == 1) in saveCameraParams, file /home/lwalter/other2/opencv_latest/samples/cpp/tutorial_code/calib3d/camera_calibration/camera_calibration.cpp, line 611
terminate called after throwing an instance of 'cv::Exception'
  what():  /home/lwalter/other2/opencv_latest/samples/cpp/tutorial_code/calib3d/camera_calibration/camera_calibration.cpp:611: error: (-215) rvecs[i].rows == 3 && rvecs[i].cols == 1 in function saveCameraParams

Aborted (core dumped)

This is using the latest source, I could try going back to a release though I'll have to track down when the useFisheye logic was added to camera_calibration.cpp. I'll run this through a debugger as well.

Oops this is the result of applying cornerSubPix to circle grid results...

The original file:

The results of findCirclesGrid with a width of 5, height of 9, and cv::CALIB_CB_ASYMMETRIC_GRID:

Should it be doing this poorly (2.4.6)?

I'm going to experiment with adaptive thresholding or some other preprocessing to see if it can improve the results.

It appears that OpenCV is writing yaml 1.0 but PyYaml only wants to read yaml 1.1 (http://pyyaml.org/#Whatimplementationshouldyouuse and http://stackoverflow.com/questions/12058510/xml-or-yml-parsing-in-opencv-with-python). Do I just need a script to reformat the yaml so it doesn't generate errors like expected alphabetic or numeric character, but found ':' on the first line %YAML:1.0, or is there a better way to go about loading the files?

I've tried cv2.cv.load(filename) but it generates and error about not finding xml tags (opencv 2.4.6 and whatever python-opencv bindings are provided in Ubuntu 13.10).

It would be good to use std::abs or fabs or the epsilon check will fail if not using namespace std.

It would be really convenient if this function worked like

  cv::distanceTransform(mask, dist, labels, distance_type, 3, DIST_LABEL_PIXEL);

and then within a double for loop on y and x

  const int label = labels.at<int>(y,x);
  const int label_x = label % width;
  const int label_y = label / width;

But instead the label is dependent on the non-zero pixels in the mask input (the label only increments when pixels are zero), so I have to manually build up a mapping between labels and pixel coordinates to find out where the labeled pixels actually are- or is there an easier way?

I'd take this to the dev zone but code.opencv.org appears to be down.

In my somewhat old git version of opencv disttransform.cpp:

   else    // it would be nice if DIST_LABEL_PIXEL was here for clarity
   {
   int k = 1;
   for( int i = 0;  i < src->rows; i++ )
   {
      const uchar* srcptr = src->data.ptr + src->step*i;
      int* labelptr = (int*)(labels->data.ptr + labels->step*i);

      for( int j = 0; j < src->cols; j++ )
        // it would be nice to get rid of 
        // this conditional and 
        // always increment labelptr, but then 
        // would the icvDistanceTransformEx_5x5_C1R
        // be messed up?
         if( srcptr[j] == 0 ) 
            labelptr[j] = k++;
      }
  }

  icvDistanceTransformEx_5x5_C1R( src->data.ptr, src->step, temp->data.i, temp->step,
                    dst->data.fl, dst->step, labels->data.i, labels->step, size, _mask );

And correspondingly does calling it with MORPH_DILATE the same as dilate()?

Here is a decoration free fullscreen (which I overlooked before) which doesn't quite answer my question but may be suitable for now:

cv::setWindowProperty("mywindow", WND_PROP_FULLSCREEN, CV_WINDOW_FULLSCREEN);

This always fullscreens on the monitor that has the menus on it in a multi monitor setup which is not ideal- I want the fullscreen output to be on a screen that is uncluttered by menus because the point of it is to be a presentation output. Changing the which monitor has the menus is a matter of dragging the menu in the Ubuntu System Settings | System Tools | Displays dialog.

Being able to remove window decoration (the bar with the minimize/maximize buttons and name, and any window border) from a highgui namedWindow would be ideal, but that doesn't seem possible.

I'm currently considering using SDL which I think SDL_NOFRAME and SDL_SetVideoMode() solves my problem, but I'm curious if there is a faster or easier way. I don't want it to be too difficult to convert the cv::Mat to the new format, and it should also be high performance (low-latency and 10s of fps framerate).

This is using Linux and X windows, and the latest 2.4 OpenCV with Qt and OpenGL built in, so making X, Qt, or OpenGL calls would be acceptable.