Pose Invariant Generic Object Recognition with Orthogonal Axis Manifolds in Linear Subspace

Abstract

This paper addresses the problem of pose invariant Generic ObjectRecognition by modeling the perceptual capability of human beings. Wepropose a novel framework using a combination of appearance and shapecues to recognize the object class and viewpoint (axis of rotation) aswell as determine its pose (angle of view). The appearance model of theobject from multiple viewpoints is captured using Linear SubspaceAnalysis techniques and is used to reduce the search space to a fewrank-ordered candidates. We have used a decision-fusion basedcombination of 2D PCA and ICA to integrate the complementary informationof classifiers and improve recognition accuracy. For matching based onshape features, we propose the use of distance transform basedcorrelation. A decision fusion using Sum Rule of 2D PCA and ICA subspaceclassifiers, and distance transform based correlation is then used toverify the correct object class and determine its viewpoint and pose.Experiments were conducted on COIL-100 and IGOIL (IITM Generic ObjectImage Library) databases which contain objects with complex appearanceand shape characteristics. IGOIL database was captured to analyze theappearance manifolds along two orthogonal axes of rotation.