Doreen Hii¹ (); ¹University of California, Irvine

The first fully autonomous, theory-driven model of 3D object reconstruction from a single 2D orthographic image, is presented. The theory on which my model is based (Pizlo & de Barros, 2021) states that the formation of the mental representation of a 3D shape is possible because the human visual system implements powerful a priori constraints such as symmetry, compactness, and planarity. My model, Bilaterally Symmetric 3D reconstructor (BiS3D), operates in two stages. In the first stage, the model reconstructs a template 3D object through the following steps: (a) extracting pairs of similar 2D contours, (b) reconstructing these pairs in 3D by applying the symmetry prior, (c) propagating the 3D information to all visible 2D edges using planarity of faces, and (d) recovering the back, occluded part of the object by mirror-reflecting one half of the reconstruction with respect to the plane of symmetry. The last step guarantees a volumetric output with all elements of a 3D combinatorial map: vertices, contours, surfaces and volume (Brun & Kropatsch, 2003). In the case of orthographic projection, symmetry and planarity priors only reconstruct a template 3D object without specifying its aspect ratio. Therefore in the second stage, the model optimizes the aspect ratio of the template object by maximizing compactness, producing a unique 3D interpretation. I tested the model in a shape constancy experiment where the task was to respond if two successive presentations of 2D images were from the same 3D polyhedral shape. Shape constancy performance of the model was evaluated by Intersection over Union of the two reconstructed 3D shapes. To demonstrate the effect of the compactness prior, I varied the degree of differences between the slants of the symmetry plane used in the two presentations. The performance of the subjects and the model was systematically affected by this manipulation.