Camera Intrinsics for camera 2 not returned by estimateCa​meraParame​ters.

Hi, can anyone explain the following?

After running the below commands:

[imagePoints,pairsUsed] = detectCircleGridPoints(CameraOneImages.Files,CameraTwoImages.Files,[9 11],PatternType="asymmetric",circleColor="white");
centerDistance = 2.2;
worldPoints = generateCircleGridPoints([9 11], centerDistance);
imageSize=[1701 1651];
params = estimateCameraParameters(imagePoints,worldPoints, 'ImageSize',imageSize);

When I check for the intrinsic matrix for camera 2, I found that "Camera Intrinsic" for camera 2 is not returned by the function estimateCameraParameters. What could be the cause for not reporting the "Camera Intrinsic" for camera 2? Everything appears to be okay since the function detectCircleGridPoints detects the centers of all the white dots in all the images and did not report any warnings and errors.

[Matt J]

We cannot run your code unless you provide imagePoints,pairsUsed in a .mat file.

[hlseck]

Hi Matt, I have provided the imagePoints and pairsUsed in the attached .mat file.

Thanks in advance.

Answer 1

[Giridharan Kumaravelu]

It looks like the estimateCameraParameters function did not have sufficient number of quality points to converge to a better solution. Because of that it did not go far away from its initial estimates for the camera parameters (which can be negative too).

Looking at the plots generated by @Matt J, it appears that the calibration images lack variety in poses and also there were only 6 images.

@hlseck Please increase the number of calibration images and also calibrate with sufficient variety as described in this doc page: https://www.mathworks.com/help/vision/ug/prepare-camera-and-capture-images.html#mw_abbcc91e-07ab-4e40-a3b6-bb4a263e8379.

Answer 2

[Matt J]

Hmmm. It seems like it might be a version-related problem. When I run your code in R2022a, I get intrinsics for both cameras:

>> params.CameraParameters1.Intrinsics

ans =

cameraIntrinsics with properties:

FocalLength: [5.8522e+05 3.4051e+05]

PrincipalPoint: [3.0379e+03 1.4885e+03]

ImageSize: [1701 1651]

RadialDistortion: [-2.0671 -2.6121e+05]

TangentialDistortion: [0 0]

Skew: 0

IntrinsicMatrix: [3×3 double]

>> params.CameraParameters2.Intrinsics

ans =

cameraIntrinsics with properties:

FocalLength: [5.6750e+05 3.0497e+05]

PrincipalPoint: [786.3843 1.0454e+03]

ImageSize: [1701 1651]

RadialDistortion: [-72.1295 1.0642]

TangentialDistortion: [0 0]

Skew: 0

IntrinsicMatrix: [3×3 double]

However, running it here in the forum (which is R2022b), generates significantly different intrinsics for camera1 and errors for camera2:

load('imagePoints_pairsUsed.mat')

centerDistance = 2.2;
worldPoints = generateCircleGridPoints([9 11], centerDistance);
imageSize=[1701 1651];
params = estimateCameraParameters(imagePoints,worldPoints, 'ImageSize',imageSize);

params.CameraParameters1.Intrinsics

ans = 
  cameraIntrinsics with properties:

             FocalLength: [5.9768e+05 3.4533e+05]
          PrincipalPoint: [2.0518e+03 1.1783e+03]
               ImageSize: [1701 1651]
        RadialDistortion: [-16.0594 -4.4515e+05]
    TangentialDistortion: [0 0]
                    Skew: 0
                       K: [3×3 double]

params.CameraParameters2.Intrinsics

Error using vision.internal.cameraIntrinsicsImpl>parseInputsSimulation
The value of 'principalPoint' is invalid. Expected principalPoint to be positive.

Error in vision.internal.cameraIntrinsicsImpl>parseInputs (line 191)
r = parseInputsSimulation(defaultParams,varargin{:});

Error in vision.internal.cameraIntrinsicsImpl (line 82)
r = parseInputs(varargin{:});

Error in cameraIntrinsics (line 67)
this = [e-mail](varargin{:});

Error in vision.internal.calibration.CameraParametersImpl/get.Intrinsics (line 502)
intrinsics = cameraIntrinsics(this.FocalLength, ...

[Matt J]

I vaguely wonder if the cameras were placed too close together (in comparison to their distance from the calibration pattern) to obtain a good calibration. The focal lengths look unrealistically large to me.

[hlseck]

Yes, the distance between camera centers is smaller than the distance from the camera to the calibration target. The ratio of their distances is about 0.3. Is there a way to constrain the focal length so as to get a more reliable calibration.

[Matt J]

0.3 doesn't seem too bad ...

No, I don't believe there is a way to constrain the parameter estimates. However, you might try varying the position of the calibration pattern a bit more among your 6 images. Currently, they all seem to give points in approximately the same positions.