The calibration grid should be place in front of the two cameras so that it is fully visible in both images. These images should then be preprocessed with the Canny operator (Edge Tool). The grid should be viewed so that the first element of the grid data file corresponds to the botton left hand vertex in both images. Illumination should not saturate the images as this can result in badly estimated (even systematically shifted) edge locations. The tile should be inclined at approximately 30 degrees to the cameras focal axis for best results (some degree of perspective distortion is necessary for good parameter estimation).
The vertex positions should be located using the "grid" process. Failiure to recover any full line of vertices across the grid signals incorrect termination, though a few missing vertices here or there can be tolerated.
Obtain an initial estimate of the calibration model (eg read in an old file). This should be either an initial guess of all parameters (in which case you can move on directly to "IP min") or a very good estimate of image centres and aspect ratios.
Use the Tsai algorithm to obtain an initial estimate of the remaining parameters.
Image plane minimisation should be attempted with a simple camera model (ie only focal length included from the model parameter choice menu). Note that when calibrating with a grid, image centres are nearly always very badly determined so these should be either switched off or constrained using a covariance matrix. On no occasion should it be necessary to include both the x and y aspect ratios in the model as either one plus the focal length parameter is sufficient to provide a full parametric description of a projective model.
After calibration the residual distributions should be examined. For good data radial distortion residuals should form a narrow band (less than 1 pixel across) down the centre of the plot. If recalibration is necessary this can proceed from the current estimate but may require the inclusion of more model parameters to obtain good results.
When a sufficiently good calibration is achieved the inverse stereo covariance matrix should be computed and written to file. Subsequent recalibration of the same fixed camera system can then be combined with this current estimate using the IPS min process. Copy the resulting calibration back to the Tina system for use in the stereo matching process as required.
The facility to use radial distortions has not yet been fully integrated into the stereo matching algorithms in Tina. The effects of this distortion can be visualised by calibrating with and without this degree of freedom in the model.