A single line drawing determines infinitely many 3D interpretations. Despite this ambiguity, human observers usually perceive a single 3D shape and the percept is veridical. The underlying perceptual mechanisms are still unknown. In this study we tested the subjects' perception of symmetrical 3D shapes, and developed a computational model of this visual ability. By comparing the subjects' percept with the shapes recovered by the computational model, we explored the underlying mechanisms of 3D shape perception. By comparing the subjects' percept with the original shapes, we explored the phenomenon of shape constancy. Method: 100 random symmetrical polyhedra were generated. Each polyhedron was presented at a randomly selected viewing orientation at one of five slants of the plane of symmetry: 15, 30, 45, 60 or 75 degrees. For each polyhedron an orthographic image was computed and hidden edges were removed. It is known that a single orthographic image of a 3D symmetrical shape determines a one-parameter family of 3D symmetrical interpretations characterized by an aspect ratio. The subjects were asked to adjust the aspect ratio of a rotating 3D shape so that this shape agreed with the 3D percept produced by a single 2D line drawing of this shape. Results: A computational model of human shape perception involves several constraints: symmetry, planarity of contours, maximum compactness (MC) and minimal surface (mS). MC maximizes the ratio V^2/S^3, whereas mS minimizes S, where V is the volume and S is the surface area of a 3D object. The subjects' percept is strongly correlated with the original shape (i.e. shape constancy is achieved) and with a model that combines MC and mS constraints. Shape constancy is reliably achieved for slants of the symmetry plane between 30 and 75 deg. The slant of 15 degrees leads to systematic errors in both the recovered and perceived shapes.