Object Recognition Using Eigenviews

We introduce a novel method to characterize the shape of objects under viewpoint variation for use in an object recognition task. Images of a collection of 3D objects captured under different viewpoint locations are used to obtain representative views (eigenviews) that encode the information in these images. Three techniques are used to extract eigenviews from a given collection of images, Principal Component Analysis, Independent Component Analysis and Nonnegative Matrix Factorization. The idea is illustrated with a collection of four synthetic objects. A Nearest Neighbor classifier is used to perform the classification of an arbitrary view of an object. The proposed system is shown to be robust to noise and partial occlusions. The effect of the number of eigenviews used is also presented. The classification results demonstrate that this system holds promise for use in object detection under variation in viewpoint location.