Rotation, Rescaling and Occlusion Invariant Object Retrieval

This paper presents a new approach for rotation, rescaling and occlusion invariant retrieval of the objects of a given database D. The objects are represented by means of many 2D views and each of them is occluded by several half-planes. The remaining visible parts (linear cuts} as well as the whole views are stored in a new database D` and described by low-level features. Given a portion R of an image, the retrieval of the most similar object is done by generating some linear cuts of R, and by comparing their descriptors with those of the elements of D`. Some heuristic rules regarding visual similarity and geometric properties of the objects in the database drive this process. In the case R is recognized as an object partially occluded, a strategy for the reconstruction of the whole shape of R is also presented. The tests carried out on synthetic and real-world datasets showed good performances both in recognition and in reconstruction accuracy.