Abstract
In a 3-D model based vision method, aspect identifier matrices can decide feature correspondences between 3-D models and an observed image by simple linear operations. This paper proposes new identifier matrices which significantly reduce aspect misidentification rate by linear(especially orthogonal)transforms. The transforms are derived from generalized eigenvalue problems. Performance of the new identifier matrices is compared with that of conventional ones. Numerical simulations on synthetic images of a pentagonal prism show that the proposed matrices reduce misidentification rate which is caused from nonlinear perspective distortion of camera.
