Cobb%MIT-OZ@MIT-MC.ARPA (04/30/84)
[Forwarded from the MIT bboard by SASW@MIT-MC.]
The Use of Perceptual Organization for Visual Recognition
DAVID LOWE
May 7, 1984 4:00PM
NE43 - 8th floor Playroom
The human visual system has the capability to spontaneously
derive groupings and structures from an image without higher-level
knowledge of its contents. This capacity for perceptual organization
is currently missing from most computer vision systems. It will be
shown that perceptual groupings can play at least three important
roles in visual recognition: 1) image segmentation, 2) direct
inference of three-space relations, and 3) indexing world knowledge
for subsequent matching. These functions are based upon the
expectation that groupings reflect actual structure of the scene
rather than accidental alignment of image elements. A number of
principles of perceptual organization will be derived from this
criterion of non-accidentalness and from the need to limit
computational complexity. The use of perceptual groupings will be
demonstrated for segmenting image curves and for the direct inference
of three-space properties from the image.
Much computer vision research has been based on the assumption
that recognition will proceed bottom-up from the image to an
intermediate 2-1/2D sketch or intrinsic image representation, and
subsequently to model-based recognition. While perceptual groupings
can contribute to this intermediate representation, they can also
provide an alternate pathway to recognition for those cases in which
there is insufficient information for deriving the 2-1/2D sketch.
Methods will be presented for using perceptual groupings to index
world knowledge and for subsequently matching three-dimensional models
directly to the image for verification. Examples will be given in
which this alternative pathway seems to be the only possible route to
recognition. A functioning real-time vision system will be described
that is based upon the direct search for the projections of 3D models
in an image.
Refreshments: 3:45PM
Host: Professor Patrick H. Winston