Look at effect of image scaling / interpolation
Comment from Matt ... I was wondering on the train - did we ever check whether computing camera calibration on 1920 x 1080 is better/worse than 1920x540?
Have re-read all the Tsai papers, he makes a big deal about even the tiniest errors in Cx,Cy parameters and the aspect ratio between dx and dy. Also, computing radial distortion must be affected.
I believe we currently calibrate on 1920x540, and just scale up the image is that right? Or more precisely, we capture 1920x540, we scale up so that square detection works, we then scale down, and then we calibrate using points effectively from a 1920x540 image, which must somewhere fold in interpolation errors from the rescaling, and may not optimise well, as you have a squashed view of the world.
Let's try and look at this as part of the SPIE paper, deadline Jan 2017