A computational model that recovers the 3D shape of an object from a single 2D retinal representation

Human beings perceive 3D shapes veridically, but the underlying mechanisms remain unknown. The problem of producing veridical shape percepts is computationally difficult because the 3D shapes have to be recovered from 2D retinal images. This paper describes a new model, based on a regularization approach, that does this very well. It uses a new simplicity principle composed of four shape constraints: viz., symmetry, planarity, maximum compactness and minimum surface. Maximum compactness and minimum surface have never been used before. The model was tested with random symmetrical polyhedra. It recovered their 3D shapes from a single randomly-chosen 2D image. Neither learning, nor depth perception, was required. The effectiveness of the maximum compactness and the minimum surface constraints were measured by how well the aspect ratio of the 3D shapes was recovered. These constraints were effective; they recovered the aspect ratio of the 3D shapes very well. Aspect ratios recovered by the model were compared to aspect ratios adjusted by four human observers. They also adjusted aspect ratios very well. In those rare cases, in which the human observers showed large errors in adjusted aspect ratios, their errors were very similar to the errors made by the model.