The perception of surface relief from random shading patterns is measu
red by having observers adjust three-dimensional local probes, the pro
jections of which are superimposed on the image. Three observers perfo
rm four settings of 91 probes on each of 14 images. These images are g
enerated by calculating the Lambertian reflectance of a random superpo
sition of elliptical Gaussian hills and valleys illuminated by a singl
e distant light source as well as by ambient light. Neither the surfac
e reflectance equation nor the light source direction is conveyed to o
ur observers in any way. Mathematically, this ''pure'' shape-from-shad
ing problem has highly non-unique solutions, Perception of a well-defi
ned, stable shape therefore implies that the ambiguity is resolved, i.
e. a gauge is fixed, We analyse the surface ambiguity or gauge freedom
which is left unconstrained by pure shading information and we invest
igate possible ways of restricting it. Statistical analysis of the cur
l component of the field of probe settings reveals that the settings a
re significantly consistent with an underlying perceived surface. In s
pite of the large theoretical ambiguity in the stimuli, the settings a
re reproducible and show considerable inter-observer agreement, Even t
he correlation of the settings with the real surfaces is surprisingly
large, If the settings are compared to the real surface normals, one f
inds a series of biases, the strongest of which is that the global sur
face slant is systematically underestimated, even in those cases where
ending occluding contours or high-contrast luminance ridges, indicati
ve of ''almost'' contours, are present in the image. Another bias then
is that the corresponding rims on the surface are seen as roughly par
allel to the picture plane.