ECVP 2011 Abstracts

Abstract

Visual tracking of non-rigid human and biological motion has received increased attention in several contexts, from replacing the hero character in movies to large scale video analytics. The conventional way to track non-rigid motion is done in a two-step process: (1) Learn a model from large amounts of training data; (2) Track in new footage the degrees of freedom of the prior model. Most recently these two steps have been merged into so-called non-rigid-structure-from-motion techniques (NRSFM), that circumvent many shortcomings of the traditional model based approach, but also introduce new challenges. We discuss both paradigms and demonstrate them on a range of problems, from high-end Hollywood productions to very high-degree-of-freedom human pose and gesture analytics applied to athletes, academics, and politicians. What do deforming shapes reveal about structure from motion Many organisms and objects deform when moving, requiring perceivers to separate shape changes from object motions. We have discovered interesting details about human perception of deforming shapes from motion cues, by using movies of rigid and flexing point-light cylinders rotating simultaneously around the depth and vertical axes in perspective: (1) Observers can discern cross-sectional shapes of flexing and rigid cylinders equally well, suggesting no advantage for structure-from-motion models using rigidity assumptions. (2) Symmetric cylinders appear asymmetric when oblique rotation axes generate asymmetric velocity profiles, highlighting the primacy of velocity patterns in shape perception. (3) Inexperienced observers are generally incapable of using motion cues to detect inflation/deflation of cylinders, but this handicap can be overcome with practice equally well for rigid and flexing objects. (4) Observers successfully classify cylinders as rigid, flexing in depth, or flexing in the image plane from combinations of motion and contour cues, but not contour cues alone. Parsing image velocity flows into kinematic differential invariants, the gradient of def is zero for rigid but non-zero for flexing cylinders, and combinations of def with curl and div classify plane and depth deformations respectively. The visual system could use these invariants to confirm rigidity and identify shape deformations. [Supported by NEI grants EY07556, EY13312] Perceiving dynamic faces A Johnston (University College London; e-mail: a.johnston@ucl.ac.uk) When faces move we can recover information about what parts of the face move together and which parts move independently. Thus motion allows us to determine how faces look on average, how they vary with respect to the average and what information it makes sense to code separately. Thus motion is critical to the representation of both moving and static faces. Principal components analysis can be used to construct a "face space" encoding a range of expressions from the movements of a single individual. Face adaptation has been used as evidence for norm-based coding of facial identity. We asked whether adaptation could alter perception in the case of a single individual's expression space. We took expressions 1 standard deviation from the mean expression along the directions of the first and second principal components. Adapting to one end of a continuum in expression space shifted the perception of faces along that axis but had no effect on the orthogonal dimension. In this experiment subjects had no experience of the target face. The results imply that adaptation does not affect the internal representation of a particular identity, rather adaptation modifies the representational structure on which the description of individual faces is based.