Abstract
A model is described that estimates object posture by using a number of ellipses. First, the ellipses are rotated and arrayed in the annular direction. Next, the ellipses that change long-and-short axis ratio are arrayed in the radial direction. Finally, a radial-ring network model is built by connecting each of these ellipses to form a concentric network. A variable template network (VTN) model is based on the same concept: an ellipsoid is rotated, enlarged and reduced, and then several a 2D appearance figure is projected onto each 2D plane surface. The VTN model is constructed by connecting these figures to form a grid. Although object postures can be estimated using this model, there are multiple figures with the same rotation angle and form in the VTN model. Similarly, since the discrete degree at the time of projection is constant, if the skewness of a 3D object is high, this can not estimate its posture accurately. The proposed model does not suffer these problems, enabling it to estimate object position and posture quickly and accurately.
