seeing in 3d

Upload: nadeem-graham

Post on 04-Apr-2018

216 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/29/2019 Seeing in 3D

    1/2

    Seeing in 3DThe world is represented via two dimensional projections onto the retina yet werecognise objects situated at unfamiliar viewpoints. There are currently two theoriesof visual object recognition:-View-point independent and view-point dependent.

    View-point independent posits the following,"On these theories, recognition consistsin the abstraction of a structural specification of the object perceived and a match of

    that specification to a memorial specification.Such theories are called viewpoint-independent because they predict that, as long as the crucial structural informationcan be abstracted from the percept, then there should be no variation in recognition

    speed and accuracy across unfamiliar rotations of the presented object (where

    rotations are either in depth or along the picture plane."According to viewpoint-independent theories, the representations that facilitate objectrecognition are specifications of the three-dimensional structure of seen objectsabstracted from the two dimensional array of the retinal image.

    Nishihara and Marr(1978) proposed an account with regards to the formatting of

    information about three-dimensional object shape in memory.They believe thatobjects are represented by a coordinate system the axes of which are centred onmajor parts of the object.Only the object is represented and not the viewers relative

    position.Each of the major parts of an object would be represented by cylinders so

    that... "a person is represented as six main cylindrical shapes.....The important pointis that objects are represented in terms of their major axes of elongation ascylindrical parts centred on those axes." The model s predicts recognition

    performance will be equivalent across different views as long as information about

    the major axes of elongation of the objects is retrievable from those views.Analternative view within the VP independent paradigm is that of Biederman and

    Lowe(1985) who posits 36 shapes,or 'geons' as he terms them,representing whatLowe calls non-accidental features of objects." According to Lowe, objects havemany aspects that remain constant even when seen from different distances and

    points of view. For instance, in many different projections of a book, the parts thatcorrespond to the books edges remain parallel. Also, symmetrical objects look

    symmetrical despite widely varying changes in distance and angle. Acoin seen fromdifferent angles may project sometimes a circular and sometimes an elliptical image

    (and sometimes a thin rectangular image) but either way, the images retain thesymmetry of the coin. Lowe defined about a half-dozen of such non-accidentalproperties of visually presented objects. Based on these non-accidental properties,Biederman devised his system of geons, examples of which include cylinders, cones,and prisms.Biederman and his colleagues predict that as long as such geon

    information is recoverable from a view of a previously seen object, different rotationsof the object should not affect recognition. Thus if, for example, you had seen a coffeemug for the first time, successive views of the mug would facilitate recognitionequally as long as both the body and handle geons of the mug were extractable from

    the retinal array. Biederman predicts that, as long as the handle is in view,rotationaldeviations from the original presentation will not affect recognition. Thus the theoryis

  • 7/29/2019 Seeing in 3D

    2/2

    a viewpoint-independent theory of visual object recognition. Many studies, however,have demonstrated viewpoint-dependent effects on recognition."

    Bartram (1974 1976), "was to find that when subjects learned a set of novelpolygonal shapes at a single rotation, reaction times for familiarunfamiliarjudgments on the shapes rotated in the picture plane,depended on the degree of

    rotation from the original orientation. Bartram (1974) found that subjects reactiontimes in a naming task decreased more rapidly across practice trials in which each of

    the eight blocks presented the object from the same view as opposed to blocks inwhich pictures of objects were seen in different views.Bartram (1976) investigatedrecognition performance across changes in viewpoint in a sequential matching

    paradigm. Subjects made samedifferent judgments on pairs of line drawings of

    objects, members of which were presented sequentially. Reaction times were faster foridentical pictures of the same object than for pictures of the same object appearing indiffering viewpoints. Jolicoeur (1985) found that naming times for line drawings ofnaturally occurring objects increased as the drawings were further from thecanonical upright views of the objects.Palmer et al. (1981) studied the recognition of

    familiar objects rotated in depth around their vertical axes. The researchersestablished independently each objects canonical view via subjects ratings of

    preferred view. In subsequent tests, they found that the subjects naming timesdecreased as the rotation of the objects away from their canonical views increased.

    Evidence of viewpoint-dependent effects in recognition has led researchers tohypothesize that recognition involves the alignment of the perceptual representationto either a single canonical view (Palmer et al. 1981), or to multiple views (Tarr and

    Blthoff 1995; Tarr and Pinker 1989). Both single and multiple view theories posit

    normalization mechanisms underlying object recognition."Another influential study is that of Shepard and Cooper 1982) who had subjects lookat simultaneously presented pairs of objects. The second member of each pair waseither the same as the first or a mirror image. Further, pair members could differ

    from each other in their rotations in depth and in the picture plane. The researchers

    found that the time it took for subjects to make samedifferent judgments increasedmonotonically with increases of rotational displacement between pair members.Shepard and colleagues took this reaction time data as evidence that subjects wererotating mental images to see if they would match the stimulus. Analogously, then,

    recognition may involve the rotation of a percept to match it against one or morecanonical views stored in memory.The evidence suggest that the dependence theory best explains the phenomena though

    studies are ongoing.