Abstract
A 3D object seen from different views forms quite different retinal images. Humans are very good at inferring 3D pose by using knowledge of projective geometry (Koch et al PNAS 2018). Everyday observations suggest that we also infer correct 3D shapes despite projective distortions, but is that true? We presented frontal views of 3 rectangular parallelepipeds lying on the ground in 16 poses each (equivalent to 16 views of one pose), and 6 observers adjusted the height of an orthogonally attached narrow cylinder to equate the physical lengths of the two limbs. The projected length of the parallelepiped changes with pose as a distorted sinusoid, but the projected length of the vertically oriented cylinder stays constant. Length estimates of the parallelepiped were close to veridical for fronto-parallel poses, 2but were seriously underestimated for poses pointing at or away from the viewer. The variation in estimated lengths correlated with projected length. The inverse of the function relating projected length to pose, gives the optimal correction factor for inferring correct physical lengths from retinal images. Observers’ correction factors were close to optimal for poses close to fronto-parallel, but seriously low for poses close to line-of-sight. Interestingly the underestimation increased with physical length of the parallelepiped. Inspection of the stimuli revealed that objects in poses away from front-parallel were seen as inclined more toward the viewer, equivalent to increases in viewing height, even with a grid on the ground plane. Increased viewing height requires smaller correction factors, so an overestimate of viewing height can explain the underestimation of object length. Since changes in perceived length of one limb with respect to the other define one class of shape distortion, these results show that shape inconstancy results despite using the correct geometric back-transform, if retinal images invoke wrong estimates of viewing height.
Acknowledgement: NIH EY13312