In most cases an image primitive will have more than one potential match associated with it. The problem is to have some criteria that identifies the correct match from the ``ghosts''. The technique used considers the neighbourhood support offered to each primitive (in a potential match). The neighbourhoods are circular with the potentially matching primitives as centre points.

The amount of neighbourhood support will be expressed as a matching strength value, the size of which suggests which potential match is the correct one. Calculating a matching strength is performed as follows,

(a)

The left primitive's neighbourhood is scanned top to bottom, right to left. Each time a primitive is found the corresponding raster in the right primitive's neighbourhood is scanned (right to left) for the next successive match. Note that the uniqueness constraint still applies here as well as the maximum edge angle difference. Before a successful match can be noted there is one last disparity gradient test.

(b)

Let the left centre primitive and the left neighbour primitive be the end points of a dipole (the same with the right image centre and neighbour primitive). If the disparity gradient calculated using these end points doesn't exceed the limit then that neighbour offers support to the centre primitive match.

(c)

Support given by the neighbour match is inversely proportional to their cyclopean distance from the centre match, this is added to the current matching strength value. Once the whole of the neighbourhood has been considered the centre match will have its final matching strength value.

[ Stage One - Find Potential Matches | Stage Three - Relaxation ]

Comments to: Sarah Price at ICBL.