Recognition and Reconstruction of Partially Occluded Objects

A new automatic system for the recognition and reconstruction of rescaled and/or rotated partially occluded objects is presented. The objects to be recognized are described by 2D views and each view is occluded by several half-planes. The whole object views and their visible parts `linear cuts` are then stored in a database. To establish if a region R of an input image represents an object possibly occluded, the system generates a set of linear cuts of R and compare them with the elements in the database. Each linear cut of R is associated to the most similar database linear cut. R is recognized as an instance of the object O if the majority of the linear cuts of R are associated to a linear cut of views of O. In the case of recognition, the system reconstructs the occluded part of $R$ and determines the scale factor and the orientation in the image plane of the recognized object view. The system has been tested on two different datasets of objects, showing good performance both in terms of recogniti on and reconstruction accuracy.