Calibration and camera orientation only with (coded) circular targets

[user]

Industrial Photogrammetry uses coded and non-coded circular targets in order to calibrate the camera and compute the orientation. Later on the dense-cloud is computed.

In the photogrammetric community it is well known, that the detection accuracy for circular targets is superior to sift or equivalent, since they can be detected with ~ 1/100px accuracy while sifts are ~ 1/5.

Therefore it would be nice to have a feature where I could choose that PhotoScan should choose only coded and non-coded circular targets for the orientation and calibration process and sift targets to generate the dense clouds.

Thank you

[Yoann Courtois]

Hello !

You should be able to calibrate your camera at first, by define a high marker accuracy (0.01pix) and a super low tie point accuracy (1 000pix).
And then, re-inject that calibration (and fix it) in a normal optimize project (marker accuracy 0.1 pix / tie point accuracy 1 pix) to go further in the workflow !

Nevertheless, I guess a well-cleaned tie point cloud would give you the possibility to get a longer polynomial distortion model (using k3-k4 & p3-p4), whereas using only targets would give you, maybe a more accurate, but a shorter polynomial distortion model (only until k2 & p2)

Regards

[user]

Hello !

You should be able to calibrate your camera at first, by define a high marker accuracy (0.01pix) and a super low tie point accuracy (1 000pix).
And then, re-inject that calibration (and fix it) in a normal optimize project (marker accuracy 0.1 pix / tie point accuracy 1 pix) to go further in the workflow !

Hello Yoann Courtois,

yes, I'm aware of this solution, but I was not sure whether PhotoScan uses the variances in order to weight the observations like it is done in weighted least squares. Thanks for pointing it out, it is indeed a quick fix.

Still, in this case the optimization process will use all points in order to create the normal equation matrix. By omitting several 100 or 1000 points it is possible to reduce the computation time in future versions.

Nevertheless, I guess a well-cleaned tie point cloud would give you the possibility to get a longer polynomial distortion model (using k3-k4 & p3-p4), whereas using only targets would give you, maybe a more accurate, but a shorter polynomial distortion model (only until k2 & p2)

At least in industrial photogrammetry there is no need for many coefficients. In most cases the lenses have a quite low distortion (unless it's a special lens like fisheye which need a special treatment anyway) and are well calibrated with 3 radial and 2 asymmetric coefficients. If necessary, one can take 2 affinity coefficients into account. For some special cases, 2 distance dependent coefficients can play a role. If we sum this up, even the worst case produces 12 unknowns (c,x0,y0,r1,r2,r3,b1,b2,c1,c2,d1,d2). Given a usual photogrammetric calibration board with roughly 1000 targets, there is enough over-determination in each image to reconstruct the distortion.

Apart from that, I'm not aware that fourth taylor-elements had any improvement when calibrated scales (and not the reprojection error) has been used to judge the accuracy.

Greetings