Correspondence-free fundamental matrix for object recognition
Gutemberg Guerra‐Filho · 2012
We propose a new method for the computation of the fundamental matrix without correspondences and a new object recognition algorithm based on the fundamental matrix as a projective invariant descriptor. The core procedure of our object recognition algorithm is the correspondence-free computation of the fundamental matrix between a curve feature in the query image and a curve feature in an object image. Our method is based on the maximization of the number of inferred correspondences between points of the two curve features that satisfy a single fundamental matrix. Based on this projective invariant descriptor for pairs of curve features, we recognize objects by clustering pairs of corresponding curve features in the space of fundamental matrices. We evaluate our correspondence-free method using synthetic data with ground truth and in the context of object recognition with real images.