Feature matching

This package implements an algorithm that finds correspond pairs from two sets of features - the "scene" and the "model". The algorithm is described in detail in "Structural matching in computer vision using probabilistic reasoning" (PhD thesis, Christmas).

The software come in two parts:

The matching algorithm matcher itself is fairly general - it can handle any set of features that can be described by attributes of the features and relations between pairs of features , where the attributes and relations are real numbers - that is, the features set is in the form of an attributed relational graph. The preprocessing utility gf2arg, that creates the graphs from gf sets, is currently restricted to handling 2-D geometric features, in particular to points and directed and undirected lines.

When matching line segments, the unknown transformation between the two feature spaces is more or less restricted to being a Euclidian transformation (i.e. unknown translation and rotation), although any prior knowledge of the Euclidian transformation can be included in the computation. Some scale mismatch can also be tolerated (typically +/- 50%). For point matching, The algorithm handles unknown translations well, but unknown orientations less well.


W.Christmas@ee.surrey.ac.uk
September 1994