We used a Kodak DC290 digital camera, capturing the images at 1200x1792.
Each line had roughly a Gaussian spread of four pixels. We had to set
the camera on exposure lock so that the images would not be too saturated
(see here for details). It is a little curious
then that a simple video camera at 720x486 could work well on the grid
as described in the Proesman paper. But they used a special slide with
very thin lines (thinner than our lcd projector).
Our grids were colored instead of white, in hopes that the color could
aid in identification and avoid a costly global correspondence algorithm.
All of our test grids can be seen here.
Furthermore, the colors could allow for exact identification along a
surface, whereas the Proesman work is modulo a translation. Being off
by a translation is not a concern given orthography. However, exact grid
location must be known to undo perspective warping or to figure out the
locations of two separate connected components.
In retrospect, it might have been easier to use a checker pattern instead
of a grid pattern (so a gradient-based detector would work). The Proesman
work uses a grid instead so that they can recover texture information
afterwards by diffusing the surrounding colors across the grids (almost like
texture synthesis).
Our specific setup took place in the dark room. We used boxes and
the Einstein pedestal for calibration, as discussed in a later section.
The projector was propped up on a box that was half leaning over a ledge
(to make it as orthographic as possible). To prevent the box from tipping,
the box was counterweighted with a digital video camera. The digital
video camera was in progressive scan mode. :)