A new theory of structure-from-motion perception

Solving the stereo correspondence problem with false matches

The stereo correspondence problem exists because false matches between the images from multiple sensors camouflage the true (veridical) matches. True matches are correspondences between image points that have the same generative source; false matches are correspondences between similar image points that have different sources. This problem of selecting true matches among false ones must be overcome by both biological and artificial stereo systems in order for them to be useful depth sensors. The proposed re-examination of this fundamental issue shows that false matches form a symmetrical pattern in the array of all possible matches, with true matches forming the axis of symmetry. The patterning of false matches can therefore be used to locate true matches and derive the depth profile of the surface that gave rise to them. This reverses the traditional strategy, which treats false matches as noise. The new approach is particularly well-suited to extract the 3-D locations and shapes of camouflaged surfaces and to work in scenes characterized by high degrees of clutter. We demonstrate that the symmetry of false-match signals can be exploited to identify surfaces in random-dot stereograms. This strategy permits novel depth computations for target detection, localization, and identification by machine-vision systems, accounts for physiological and psychophysical findings that are otherwise puzzling and makes possible new ways for combining stereo and motion signals.

Discerning nonrigid 3D shapes from motion cues

Many organisms and objects deform nonrigidly when moving, requiring perceivers to separate shape changes from object motions. Surprisingly, the abilities of observers to correctly infer nonrigid volumetric shapes from motion cues have not been measured, and structure from motion models predominantly use variants of rigidity assumptions. We show that observers are equally sensitive at discriminating cross-sections of flexing and rigid cylinders based on motion cues, when the cylinders are rotated simultaneously around the vertical and depth axes. A computational model based on motion perspective (i.e., assuming perceived depth is inversely proportional to local velocity) predicted the psychometric curves better than shape from motion factorization models using shape or trajectory basis functions. Asymmetric percepts of symmetric cylinders, arising because of asymmetric velocity profiles, provided additional evidence for the dominant role of relative velocity in shape perception. Finally, we show that inexperienced observers are generally incapable of using motion cues to detect inflation/deflation of rigid and flexing cylinders, but this handicap can be overcome with practice for both nonrigid and rigid shapes. The empirical and computational results of this study argue against the use of rigidity assumptions in extracting 3D shape from motion and for the primacy of motion deformations computed from motion shears.