Model-Based Industrial Computer Vision

Model-based computer vision refers to the use of symbolic or geometric models for object recognition, pose estimation, and other tasks. Our most recent work in this area developed a new technique called relational indexing in which a 3D object is represented by its major views, and each view is represented by a relational structure. The primitives used in the relational structure are high-level features such as ellipses, parallel pairs of lines, and coaxial arcs. The relational structure is a set of two-graphs, two-node graphs representing two primitives and the relationships between them. The relational indexing process takes an input image, extracts its features, converts them to two-graphs, and uses them to index into the database of models and vote for the view classes of those objects that have similar structures. A verification procedure using pose estimation from points, lines, and ellipses is used to decide on the correct (model,view class) pair.