Ed Gamble
Simplifying Discontinuity Detection
with an Eye on Recognition
Abstract:The essence of our approach is to address the important problem of discontinuity
detection within the context of the overall visual recognition problem. Awareness of
the capabilities and expectations embodied in recognition algorithms and of the dominant
noise process in the surface properties computed by some early vision algorithms
greatly simplifies the detection of discontinuity. In particular, we describe the characteristic
“displacement” errors, that stereo and optical-flow algorithms produce near
object boundaries and we suggest that the detected discontinuities, in light of these
errors, must be restricted to a subset of intensity edges. This restriction simplifies
discontinuity detection and is valid under certain assumptions which we describe.
We have detected discontinuities in depth and in the magnitude of optical flow
for a variety of natural images by combining intensity edges and surface property
data computed with early vision algorithms. The integration of surface properties
is formulated as an optimization problem derived from a Markov random field. A
massively parallel, stochastic relaxation algorithm for solution of these optimization
problems is described.