Relational model construction and 3D object recognition from single 2D monochromatic image

摘要

This paper reports a method for automatically constructing a relational model of a rigid 3D object to represent view-independent relations among its component parts, and of such using a model to recognize the object from single monochromatic images. The relational model is a graph associated with procedural constraints. It is constructed by a statistical analysis of images generated by projecting a CAD model of the object from a set of viewpoints on the Gaussian viewsphere. Object recognition is achieved by a hypothesis-and-verification process. Extended hypotheses are generated by aggregating image features which satisfy the view-independent constraints. These hypotheses are then verified by projective inversion and 3D grouping to achieve object recognition. The approach is illustrated by means of the recognition of a hatchback model car. The method can readily be adapted to the recognition of any object defined geometrically.