Accurate pose estimation of rigid body by manifold embedding

Most researchers believe that multi-view images taken from the same object lie on a low dimensional manifold. Based on this observation, this paper mainly focuses on answering the question: how accurately rigid object pose can be estimated by manifold embedding? Firstly, a new manifold embedding method was proposed which owns the property of preserving relative position on the low-dimensional manifold. Based on the method, a framework of pose estimation was suggested and it can retrieve high precision pose information in the noise-free case. However, noises and disturbances are unavoidable in real applications. We proved it by experiments that noises and disturbances lead to deformation of manifold, which will lead to depressed pose estimation accuracy. To tackle the problem above, an improved framework was proposed in which image segmentation techniques were adopted to improve the robustness to noises and disturbances. Finally, the improved framework shows good performance on both Gaussian white noise and illumination direction, because the effect of them has been removed by image segmentation.