Appearance Manifold with Covariance Matrix for 3-D Object Recognition

The authors present a robust 3-D object recognition system for recognizing noisy images. Since a recognition system usually deals with objects taken from various viewpoints, their appearance will vary from one viewpoint to another. Generally, the appearance of an object changes along with the changes of image conditions, and so does its position in the eigenspace. Such changes may cause an inaccurate recognition of an object. In this paper, we propose a novel object recognition method where covariance matrix calculation is embedded in parameterized appearance manifold. The appearance manifold will capture object characteristics along the pose rotations where the covariance matrix calculation will give the sample distribution information. Specifically, we propose the Appearance Manifold with Constant Covariance matrix (AMCC) and Appearance Manifold with View-dependent Covariance matrix (AMVC) methods. Experimental results showed that our approach could enhance the recognition performance, as well as perform robust recognition of 3-D objects under varying viewpoints and translation effects.