CameraPinholeDistor::normImageToPixelCoordinates is not implemented

[GoogleCodeExporter]

As this inverse transformation is not trivial, I prefer to wait Vincent's 
advises to start the implementation.

Original issue reported on code.google.com by thibault...@gmail.com on 31 May 2011 at 4:15

[GoogleCodeExporter]

Hey Thibault

I might be able to help here - I spent a while working with distortion models 
during the "calibration phase" of my thesis.

"CameraPinhole::normImageToPixelCoordinates()" will convert from undistorted 
normalized coordinates to undistorted pixel co-ordinates, right?

And you want "CameraPinholeDistor::normImageToPixelCoordinates()" to convert 
from the undistorted, normalized coordinates to the DISTORTED pixel 
co-ordinates? So you are missing the step that applies a distortion model to a 
set of normalized image co-ordinates?

In that case, I have some code that might help. Obviously it will depend on the 
particular distortion model, but I implemented a solution for the rational 
model in OpenCV and it was quite simple, and I'm sure it can be adjusted to 
suit more models (such as ones using 5 or less coefficients).

Do you want me to try and implement it? Here is a bit of a sample of my old 
code:

------------------------------------------------------------

fx0 = cameraMatrix.at<double>(0, 0);
fy0 = cameraMatrix.at<double>(1, 1);
ifx0 = 1./fx0;
ify0 = 1./fy0;
cx0 = cameraMatrix.at<double>(0, 2);
cy0 = cameraMatrix.at<double>(1, 2);
// and k is an array containing the distortion model coefficients for a 
rational model

// x & y are the normalized, undistorted co-ordinates
// Determine radial and tangential distances/factors
r2 = x*x + y*y;
icdist = (1 + ((k[7]*r2 + k[6])*r2 + k[5])*r2)/(1 + ((k[4]*r2 + k[1])*r2 + 
k[0])*r2);
deltaX = 2*k[2]*x*y + k[3]*(r2 + 2*x*x);
deltaY = k[2]*(r2 + 2*y*y) + 2*k[3]*x*y;

// x_new and y_new are the re-distorted co-ordinates, but are still normalized
x_new = (x0/icdist) + deltaX;
y_new = (y0/icdist) + deltaY;

// x2 and y2 are now denormalized as well as redistorted, so should be 
equivalent to the positions on the original image
x2 = (x1/ifx0) + cx0;
y2 = (y1/ify0) + cy0;

------------------------------------------------------------

I'm happy to help with this front-end stuff (image I/O, calibration, features) 
when I can because I'm probably more knowledgable with this than I am with the 
actual SFM algorithms! 

Original comment by stephen....@gmail.com on 3 Aug 2011 at 7:43

• Added labels: ****
• Removed labels: ****

[GoogleCodeExporter]

this would be very useful!!!
You can make changes and test via the testcamera.cpp tutorial (named "Learn 
what you can do with a pinhole distorded camera") if the 
distortion/undistortion work well. If it's the case, you can commit ;)

Original comment by thibault...@gmail.com on 12 Aug 2011 at 11:54

• Added labels: ****
• Removed labels: ****

[GoogleCodeExporter]

Hey Thibault

OK, I've added in some re-distortion and the changes seem to work (although 
there are slight floating-point inaccuracies in the results of Test 8 Option 
1). I've committed it too (clumsily - I left out the log message accidentally 
:P). So once you've verified that the changes work feel free to close this 
issue.

Steve

Original comment by stephen....@gmail.com on 17 Aug 2011 at 2:55

• Added labels: ****
• Removed labels: ****

[GoogleCodeExporter]

Great work! This issue is now closed!

Original comment by thibault...@gmail.com on 18 Aug 2011 at 10:48

• Changed state: Done
• Added labels: ****
• Removed labels: ****