Description
TitleStructure from motion without projective consistency
Date Created2016
Other Date2016-05 (degree)
Extent1 online resource (vi, 38 p. : ill.)
DescriptionStructure-from-motion (SFM) studies have shown that people are good at perceiving 3D structure in dynamic dot displays consistent with rigid object rotation. However, observers can perceive volumetric structure even when image motion is inconsistent with rigid rotation. As an extreme case, in dynamic figure-ground displays containing textural motion, observers perceive one set of regions as rotating in 3D, despite constant dot speed everywhere (projectively inconsistent with 3D rotation; Froyen et al. 2013, JOV; Tanrikulu et al., 2013, 2014, VSS). It is unclear, however, to what extent this extreme “tolerance” is due to the figure-ground competition in those displays. Here we used standard SFM displays, depicting a single object in isolation, and manipulated the discrepancy of image motion from 3D rigid rotation. We started not with 3D objects, but with a 2D velocity field within a vertically oriented ellipse. For an ellipsoid rotating about its principal axis, its orthographic-projected speed profile is a cosine function along each orthogonal “rib”. We manipulated the proportion α of cosine speed vs. constant speed (range 0-1), and the motion direction θ relative to the orthogonal ribs (range 0-60°). In experiment 1, observers used a 7-point scale to rate the degree to which the display depicted a volumetric object. In experiment 2, observers adjusted the depth to match the SFM displays. Both experiments shows that the volumetric percepts increased significantly with α, and were surprisingly tolerant to deviations from the projectively correct α=1, θ=0. For α, volumetric ratings increased between 0-0.6 but plateaued beyond 0.6. The effect of θ was surprisingly small, with even θ=45° yielding high volumetric ratings. Then we applied a rigidity-based computer vision model to our displays, and compared the model prediction with observers’ data. The comparison shows that the rigidity-based model does not predict human’s volumetric percepts correctly. In addition, the predicted motion from the model has large deviations from the display itself in terms of rotation axis. Thus even in standard SFM displays, the 3D percept was surprisingly tolerant to discrepancies from projectively correct rigid object motion. These results argue for a more nuanced view of 3D interpretation in which strict projective consistency plays a less prominent role than in conventional SFM accounts.
NotePh.D.
NoteIncludes bibliographical references
Noteby Xiaoli He
Genretheses, ETD doctoral
Languageeng
CollectionGraduate School - New Brunswick Electronic Theses and Dissertations
Organization NameRutgers, The State University of New Jersey
RightsThe author owns the copyright to this work.