Observers judged the slants and tilts of numerous regions within shaded images of ellipsoid surfaces that varied in shape, orientation, surface reflectance, and direction of illumination. The perceived three-dimensional structure of each surface was calculated from these judgments. Much of the error in observers' responses resulted from a tendency to perceive surfaces whose axes were aligned with the display screen. The presence of specular highlights or cast shadows, in contrast, had no effect on performance. The results of the experiment indicate that several assumptions of certain formal models for perception of shape from shading are not psychologically valid. The most notable of these assumptions are that the visual system initially assumes that all surfaces have Lambertian reflectance and that illuminant direction must be known before shape detection can proceed. These assumptions are often accompanied by a third assumption that surface orientation is detected locally, and global shape determined by smoothing over local surface orientation estimates. The present experiment indicates that an alternative approach offered by Koenderink and van Doorn may be more psychologically accurate, as it avoids all three assumptions.