Each object in the dataset is illuminated under 20 different directional lightings (move the mouse on the input images to see different illuminations), which are calibrated with two chrome spheres. The lighting strength is estimated by a simple normalization on image intensities (99 percentile) followed by a nonlinear optimization. The albedo and normal vectors of the object are solved with a least squares system, and the depth map is integrated with the Frankot-Chellappa algorithm .
The reconstruction error is measured by re-rendering the estimated normal map into a shading image and comparing that with the actual captured one. We show the median error below, which is usually below 0.01 (with intensities scale in [0, 1])
The data as well as the code for normal and surface reconstruction can be downloaded here.
 Robert Frankot and Rama Chellappa. "A method for enforcing integrability in shape from shading algorithms." Pattern Analysis and Machine Intelligence, IEEE Transactions on 10.4 (1988): 439-451.