E.B. Gamble, Jr.
Enhanced Discontinuity Detection
from Postulated Discontinuities
Abstract:We describe a discontinuity detector that integrates visual information to
suppress the pervasive noise in surface property data. The discontinuity detector
is based on the notion of 'constrained kernels' and 'postulated discontinuities.'
The kernels arise as solutions to the diffusion equation in the presence
of local, static boundary conditions provided by the postulated discontinuities.
The resulting kernels smooth the surface property data; they are Gaussian-like
except near the postulated discontinuities. Unlike most visual integration
schemes, our detector does not suffer from problems regarding computability
and parameter specification; it is fast and accurate. In addition, the detector
eliminates the pervasive ‘displacement errors’ in surface property data. We
describe these displacement errors, integrate intensity edges (as postulated discontinuities)
with stereo depth and optical-flow to compute depth and motion
discontinuities, and compare our detector to other discontinuity detection approaches.
This is a reformated version of a paper submitted to the Second European
Conference on Computer Vision in Santa Margherita Ligure, Italy. The original
submission to ECCV'92 was dated 8 October 1991.