| 3D point correspondence and pose normalization are two fundamental problems that should be solved before statistical shape models and 2D/3D shape analysis can be performed. In this work, we focus on these two fundamental problems and we want to solve them with a principled framework. We adopt the minimum description length principle, a formalization of Occam's Razor in which the best hypothesis for a given set of data is the one that leads to the best compression of the data. Specifically, two MDL-based approaches are proposed for automatically solving point correspondence and groupwise pose normalization. Exploiting nonlinear properties in point correspondences to improve results is the main motivation of the proposed point correspondence work. For pose normalization, the problem of interest is groupwise pose normalization: how to jointly pose normalize a group of 3D shape models of the same class. The groupwise pose normalization problem is formulated as an optimization problem, and a gradient descent approach is proposed to solve it. The use of landmark information is also studied in solving groupwise pose normalization.;Although the proposed methods are general, they are tested with medical applications in this work. The proposed point correspondence method is tested with different sets of 3D shapes of organs of the body, while the groupwise pose normalization method is tested for 3D craniofacial image analysis The experimental results show the proposed methods perform better than the state of the art methods to which they are compared. |
