Most research on the perception of 3-D objects in 3-D scenes has used either simulated scenes or real-world scenes. Simulated scenes provide good control over stimulus characteristics but may lack the abundance of information sources available in natural vision. Real-world scenes provide this information but it is difficult to control all attributes of objects in such scenes. In this research we introduce a hybrid method in which a simulated object—or more precisely a region on the ground plane—was superimposed on a real scene using computer graphics. Our purpose was to determine the effect of information for scene position on the perceived shape of a region in a 3-D scene, when the projected shape of the region is kept constant. The scene was produced using 50 digital photographs taken at fixed positions along a track, based on a sinusoidal function, to simulate back and forth horizontal head motion. A computer-generated L-shaped region was placed on the ground plane at one of four scene positions (distances from the observer). The projected image of the region was identical at each scene position. The region thus represented 3-D regions of different sizes and shapes, depending on the scene position. The observer's task was to adjust the horizontal leg of the L until it appeared to match the depth leg. Judged width to depth ratios increased with scene position consistent with projective geometry, but the ratios were smaller than those predicted from the geometry. This is consistent with the perception of a less slanted scene than the simulated scene. We calculated the surface slant required to match the judged ratio at each scene position and found that this slant varied with scene position. This suggests that the shape judgments for regions on the ground plane were based on local slants that varied with the simulated distance of the observer, or, equivalently, that these judgments were consistent with a curved ground plane.
Supported by NIH Grant 1R01EY12437.