Pose Estimation Under Anisotropic Scaling with Automatic Initialization

In this paper we investigate solutions to the problem of pose estimation for a class of rigid objects parameterized by anisotropic scaling (variable aspect ratios). We describe extensions of two known pose estimation algorithms to this scenario, that allow them to estimate aspect, in addition to rotation and translation. Our first contribution is a fully automatic example-based initialization scheme. We show that this scheme benefits iterative pose estimation methods across the board, dramatically improving convergence to accurate results. Our second contribution is a pose estimation algorithm called Scaled Exterior Orientation (SEO) that handles anisotropic scaling. In experiments on two object classes we show that combined with the proposed automatic initialization, SEO achieves best results in low noise conditions, and is on par with other algorithms under severe noise.