A Taj Mahal model is 'textured' by merging outputs of shader lamps.

The underlying physical model is wooden and spray painted white to create neutral surface. It is approximately positioned in front of the projectors.

Areas illuminated by first projector. 
The images are rendered using traditional graphics pipeline

Areas illuminated by second projector. Since the prespective projection parameters are computed by calibrating the projectors with respect to the physical model, the overlapping projected images are registered. The blending (merging) is achieved via modified feathering algorithm described in the paper.

Virtual Shadows (due to Sun).
Since we have complete 3D representation of the physical objects and the projectors, we can create almost any view-dependent or view-independent effects.

Closeup of 'textured' physical model (click to see bigger image).
The neutral surface appears to have detailed features. With precise alignment achieved with 3D calibration (and off-line authoring), details are 'painted' on appropriate geometrical surfaces. Note that the the projected image is a simple 2D image computed with traditional rendering process.

Large gif animation file
 
 

Registering images computed with 3D perspective rendering is achieved by first computing the internal and external parameters of the data projector. The calibration involves finding which pixels of the projector (center of cross-hair) illuminate known 3D locations on the model. (We move the computer mouse to translate the cross-hair in projector framebuffer.)

The setup with wooden model of TajMahal illuminated with projectors. The optical tracker, seen mounted on the tripod, provides the user location whichwhich is used for view-dependent effects such as specular highlights.