A novel approach to nearest neighbour search in high-dimensional spaces for 3D object recognition

This paper presents a new technique for representing shape information of 3D objects, together with the realisation of a 3D object recognition system that uses exclusively view-based information for object pose retrieval. During training, the system acquires two-dimensional views of 3D objects and automatically generates a model database built upon a local shape description of the single object views. During recognition, a two-dimensional view of the scene is matched against the model views and the objects present in the scene are recognised and localised. In order to cope with the large amount of information which is originally extracted from the model views, an adaptive technique for multi-dimensional data reduction is employed. Such a technique tales into consideration individual and intrinsic object characteristics making the amount of computation both in learning and in recognition, considerably smaller. This is achieved by adopting a new approach to nearest neighbour search in high-dimensional spaces applicable to feature vectors whose distribution follows distinct low-dimensional paths with respect to the original space dimensionality.