Stereo calibration baseline in meters

asked 2013-12-05 07:05:12 -0600

bpp
96 ●1 ●8

updated 2013-12-06 15:40:15 -0600

Hi, I calibrated my stereo webcam with a chessboard and:

stereoCalibrate(object_points, imagePoints1, imagePoints2, CM1, D1, CM2, D2, img1.size(), R, T, E, F, cvTermCriteria(CV_TERMCRIT_ITER+CV_TERMCRIT_EPS, 100, 1e-5), CV_CALIB_SAME_FOCAL_LENGTH | CV_CALIB_ZERO_TANGENT_DIST);

I need to know the baseline of the stereo camera in meters, so I look inside the T matrix. The first element of the T matrix is -1.6952669833501108e+00, the problem is that when I measure physically the distance between sensors with a ruler is approximately 4 cm, that is 0.04 m, what is the metrics of the elements in the T matrix? Therefore, I don't know the physical size of the pixels, because I don't know the sensor used.

%YAML:1.0
CM1: !!opencv-matrix
   rows: 3
   cols: 3
   dt: d
   data: [ 6.7035296112450442e+02, 0., 3.0366775716427702e+02, 0.,
       6.7334702239218382e+02, 2.3446212435423504e+02, 0., 0., 1. ]
CM2: !!opencv-matrix
   rows: 3
   cols: 3
   dt: d
   data: [ 6.7035296112450442e+02, 0., 3.0235229497623436e+02, 0.,
       6.7334702239218382e+02, 2.3428252297251004e+02, 0., 0., 1. ]
D1: !!opencv-matrix
   rows: 1
   cols: 5
   dt: d
   data: [ -2.7243204402554655e-01, 6.7461566940160000e-01, 0., 0.,
       -1.1338340048806972e+00 ]
D2: !!opencv-matrix
   rows: 1
   cols: 5
   dt: d
   data: [ -2.9489669516506523e-01, 8.7704813783022018e-01, 0., 0.,
       -1.7809117056153765e+00 ]
R: !!opencv-matrix
   rows: 3
   cols: 3
   dt: d
   data: [ 9.9994998218891662e-01, 1.8032584564876427e-03,
       -9.8377527578461001e-03, -1.7978260373390045e-03,
       9.9999822653927839e-01, 5.6101678905229378e-04,
       9.8387469692470843e-03, -5.4330216016358439e-04,
       9.9995145076190473e-01 ]
T: !!opencv-matrix
   rows: 3
   cols: 1
   dt: d
   data: [ -1.6952669833501108e+00, -8.3183862355057793e-04,
       1.6553031717676806e-02 ]
E: !!opencv-matrix
   rows: 3
   cols: 3
   dt: d
   data: [ 2.1575221682576832e-05, -1.6552550421804146e-02,
       -8.4108476712251815e-04, 3.3231506665764049e-02,
       -8.9119281968270958e-04, 1.6950218347962700e+00,
       3.8795921397114311e-03, -1.6952624768406708e+00,
       -9.5925666229836515e-04 ]
F: !!opencv-matrix
   rows: 3
   cols: 3
   dt: d
   data: [ 6.5486525518936608e-08, -5.0017984753281888e-05,
       9.9960823121244876e-03, 1.0041793870914565e-04,
       -2.6810044647830857e-06, 3.4036589992213000e+00,
       -1.5652165011118380e-02, -3.4182603874751876e+00,
       9.9999999999999989e-01 ]
R1: !!opencv-matrix
   rows: 3
   cols: 3
   dt: d
   data: [ 9.9980525157248468e-01, 2.2991356175347437e-03,
       -1.9600329167972070e-02, -2.2937864080306968e-03,
       9.9999732564029009e-01, 2.9539157402834107e-04,
       1.9600955894930411e-02, -2.5037507834536546e-04,
       9.9980785145963180e-01 ]
R2: !!opencv-matrix
   rows: 3
   cols: 3
   dt: d
   data: [ 9.9995221263036305e-01, 4.9065951283155365e-04,
       -9.7637958235422401e-03, -4.9332403320461707e-04,
       9.9999984173287226e-01, -2.7049145825487527e-04,
       9.7636615590471869e-03, 2.7529524731464261e-04,
       9.9995229642492811e-01 ]
P1: !!opencv-matrix
   rows: 3
   cols: 4
   dt: d
   data: [ 6.0327398477959753e+02, 0., 3.0294886016845703e+02, 0., 0.,
       6.0327398477959753e+02, 2.3173668479919434e+02, 0., 0., 0., 1.,
       0. ]
P2: !!opencv-matrix
   rows: 3
   cols: 4
   dt: d
   data: [ 6.0327398477959753e+02, 0., 3.0294886016845703e+02,
       -1.0227593432896966e+03, 0., 6.0327398477959753e+02,
       2.3173668479919434e+02, 0., 0., 0., 1., 0. ]
Q: !!opencv-matrix
   rows: 4
   cols: 4
   dt: d
   data: [ 1., 0., 0., -3.0294886016845703e+02, 0., 1., 0.,
       -2.3173668479919434e+02, 0., 0., 0., 6.0327398477959753e+02, 0.,
       0., 5.8984939980032047e-01, 0. ]

I tried to calibrate another stereo rig with baseline of about 8.5 cm:

image description

in this case the baseline computed is -3.5152250398995970e+00:

T: !!opencv-matrix rows: 3 cols: 1 dt: d data: [ -3.5152250398995970e+00, -3.3355285984050492e-02, -1.1137642414481164e-01 ] how is it possible?

edit retag flag offensive close merge delete

Comments

Hi, I got the similar Q as you got. my last two elementsi are 8.5807419247084848e-002, and 0. the baseline is exactly what I have. But one problem bothering me is the 0 --- When I use disparity map and Q matrix to do reprojectImageTo3d, the point cloud it creates looks like just a pyramid where most points gathers at top and I can see nothing from that point cloud. Have you met these situation before? thanks! I've been stuck on this problem so long...

bennygato ( 2014-07-14 15:36:07 -0600 )edit

add a comment

2 answers

Sort by » oldest newest most voted

answered 2013-12-08 14:53:02 -0600

jensenb
730 ●11 ●18

The unit used in the translation vector (T matrix) is the same unit used when you specify the object coordinate points. Without knowing more about your calibration setup, its hard to say why you don't see meaningful values. Some things to consider would be:

Make sure your object point coordinates are correct and match up to the detected image points
Verify the reproduction error. The error should be below one pixel is the calibration is good
Make sure the left and right image pairs correspond. Also make the board was detected in each image with the same orientation. With symmetrical boards it can sometimes happen that the detected board layout flips between the left and right view.
Ensure that your image capture of both cameras is synchronised. This is more challenging than it initially seems. If your calibration scene is not static, i.e. your are manually holding the board or the cameras, and you are not having synchronised capture you can easily get motion between the two corresponding images. This can greatly distort your extrinsic calibration results.

edit flag offensive delete link

Comments

Perfect, now I understand that I should multiply that value for 2.4 cm, that is the size of the chessboard elements. Infact -3.5152250398995970 cm * 2.4 cm= 8.4365401 cm, that seems to be coherent with the ruler's measurement of about 8.5 cm. For clarity this is the code where I specify the object coordinate points:

int numBoards = 30;
int board_w = 9;
int board_h = 6;
Size board_sz = Size(board_w, board_h);
int board_n = board_w*board_h;
vector&lt;vector&lt;Point3f&gt; &gt; object_points;
vector&lt;vector&lt;Point2f&gt; &gt; imagePoints1, imagePoints2;
vector&lt;Point2f&gt; corners1, corners2;
vector&lt;Point3f&gt; obj;
for (int j=0; j&lt;board_n; j++)
{
    obj.push_back(Point3f(j/board_w, j%board_w, 0.0f));
}

Thank you so much for your useful and complete answer.

bpp ( 2013-12-11 08:44:53 -0600 )edit

My next step will be to synchronize in hardware the two cameras so the stereo system will works with non static scene.

bpp ( 2013-12-11 08:52:23 -0600 )edit

add a comment

answered 2013-12-11 10:36:22 -0600

bpp
96 ●1 ●8

updated 2013-12-11 10:40:50 -0600

Is the focal length in the camera matrix in pixels? Since I have 2 camera matrices obviously with different intrinsic parameters what camera matrix can I use as reference the first one or the second one?

edit flag offensive delete link

Comments

yes the focal length is in pixels. It is customary to use the left camera frame as base frame. That is what opencv does when it calculates the translation and rotation during stereo calibration.

For clarity it is best to either open new questions, or post comments on the original question. You should not create answers with more questions.

jensenb ( 2013-12-11 13:55:28 -0600 )edit

Ok, thank you jensenb.

bpp ( 2013-12-14 09:52:15 -0600 )edit

add a comment

Stereo calibration baseline in meters

Comments

2 answers

Comments

Comments

Links

Question Tools

Stats

Related questions

Stereo calibration baseline in meters edit

Comments

2 answers

Comments

Comments

Links

Question Tools

Stats

Related questions

Stereo calibration baseline in meters