Establishing low dimensional manifolds for 3D object pose estimation

We propose a novel solution to the problem of 3D object pose estimation problem that is based on an efficient representation and feature extraction technique. We build a part-based architecture that takes into account both appearance-based characteristics of targets along with their geometrical attributes. This bunch-based structure encompasses an image feature extraction procedure accompanied by a clustering scheme over the abstracted key-points. In a follow-up step, these clusters are considered to establish representative manifolds capable of distinguishing similar poses of different objects into the corresponding classes. We form low dimensional manifolds by incorporating sophisticated operations over the members (clusters) of the extracted part-based architecture. An accurate estimation of the pose of a target is provided by a neural network-based solution that entails a novel input-output space targeting method. The performance of our method is comparatively studied against other related works that provide solution to the 3D object pose estimation and that are based on a) manifold modeling, b) object part-based representation and c) conventional dimensionality reduction frameworks. Experimental results justify our theoretical claims and provide evidence of low generalization error when estimating the 3D pose of objects, with the best performance achieved when employing the Radial Basis Functions kernel.