Two-view Geometry, Symmetry, and Object Perception

Aaron Michaux, Purdue University


The human vision system almost always parses visual scenes into objects, and recovers those objects accurately, forming veridical 3D perceptions. Veridical means "coinciding with reality", or seeing what is really "out there" in the world. Research is perceptual psychology has demonstrated the importance of priors in human vision, and 3D mirror symmetry prior is the most important prior. However, there has been comparatively little computer vision research in 3D mirror symmetry, despite the acknowledged importance of priors. In this thesis a principled approach is developed to the study of computer vision problems using priors, and applied to figure/ground organization (FGO), and the recovery of 3D objects from camera images. Before the new theoretical and experimental results are presented, the relevant geometrical and computational tools are reviewed. It is argued that symmetry is an informative prior in both FGO and 3D recovery. 3D symmetry aids FGO because (i) many individual objects exhibit symmetry, and (ii) configurations of unrelated objects are rarely symmetrical. Using K-means as a baseline, this approach was tested on a corpus of 180 image pairs. Symmetry based FGO outperforms the baseline in almost all cases. With respect to 3D object recovery, 3D symmetry constrains the 3D interpretation of the object image, and provides an accurate method to locate object points in R3. 3D object recovery was tested on a corpus of 89 image pairs. The average 3D reconstruction error is reported. Furthermore, animations are presented showing 3D reconstructions. Finally, two simulations were run that (i) compare the accuracy of 3D recovery against Triangulation, and (ii) estimate the expected precision and accuracy of 3D recovery.




Pizlo, Purdue University.

Subject Area

Computer Engineering

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server