The reliability of in-depth curvature judgements for linear elements was studied with stereograms that contained two linear targets and a background representing a hemisphere. The targets were arcs facing to the left or to the right, like parentheses. Some formed binocular pairs with (type 1) or without (type 2) in-depth curvature. The others were monocular (type 3). The hemisphere in the background was generated by a random curve (Ninio, 1981 Perception 10 403 – 410); it was either concave (hollow) or convex. The arcs had their binocular centre in the plane of the centre of the hemisphere. Each stereogram contained a type 1, and either a type 2 or a type 3 target. Subjects had to judge the hemisphere curvature, then the in-depth curvature of the targets in 32 different stereograms covering all curvature combinations. There were about 15% errors on type 1 targets, and 80% of these occurred when both the hemisphere and the target were convex, the target being perceived as concave, by transparency through the hemisphere. There were also about 15% errors on type 2 targets, but spread among all situations, the trend being to perceive them as slightly concave. The monocular stimuli (type 3) were judged to be frontoparallel in 70% of the cases. Otherwise, there was no directional bias except for monocular arcs on the nasal side, in conjunction with a concave background. Then, the perceived in-depth curvature was in the ‘generic’ direction predicted by associating the monocular arc in one image with a straight vertical segment in the other image.
