• What are the issues in matching/grouping algorithms using more than two
images? Is simultaneous matching really tractable, or will combinations of
pairwise matching be the norm?
This question seems to be somewhat adrift of the core grouping is
sue, but it is certainly an interesting point to ponder. I can imagine
matching algorithms analogous, in some aspects, to the bottom-
up (preattentive) phase of our grouping system. Graphs would be
constructed in which nodes represent features to be matched and
arcs represent matching compatibility by some appropriate measure.
One feature of this measure, naturally, is that the two nodes must
represent features in two different images. Cliques in this graph
w r ould then suggest matched features over some set of the images.
Of course, not all features may be visible or extracted from all im
ages, so the cliques do not have to contain as many nodes as images;
the strongest ones will, however. In this scenario, a pairwise mea
sure of compatibility would be used to build the links in the graph,
but the decision process, couched in terms of clique extraction and
analysis, would be carried out using information gleaned over the
complete image set. Is this, then, truly simultaneous matching?
Perhaps not strictly, but arguably so.
To move to true simultaneous matching over more than tw r o images
(and the need to do so is not clear to me) would require the devel
opment of N-ary match quality measures. Actually applying such
N-ary predicates over large numbers of primitives (candidate match
features) drawn from large (more than two) sets of images would
induce a considerable combinatorial burden. Only by reducing the
feature count by increasing their complexity (grouping) could this
explosion conceivably (and only conceivably) be held in check. Then
one must accept the computational cost of perceptual organization -
perhaps a desirable tradeoff given a capability such as ours. And, of
course, w r e all recognize that more complex features are less ambigu
ous, but harder to match in complicated scenes wherein portions
of the feature may be occluded, distorted, or cropped from one or
more of the other images. So we would tend to argue that true
simultaneous matching is not likely to become tractable.
Boyer - 3