next up previous
Next: CHARACTERIZING POINTS OF INTEREST Up: Differential Invariants for Previous: Differential Invariants for

INTRODUCTION.

This work addresses the problem of matching stereoscopic uncalibrated color images with the aim of computing synthetic images having intermediate points of view. The method described here is used to recover the epipolar geometry in an uncalibrated case. We revisit iconic methods for matching point primitives with local invariants in gray level images and we show that color information can robustly improve matching. Color information provides more than the multiplication of the gray level features, it provides two new invariants which characterize strongly the underlying signal. This characterization is invariant to Euclidean transformation, and we introduce here a new method to make it invariant to affine changes of image intensity. We show that matching color images can be robustly achieved on images differing from these transformations using only the eight first order local invariants (two invariants in the grey level case).

At section 2, we define a new matching method to characterize points of interest in an image. We estimate a set of local invariants which are not dependent of possible translation, rotation and change of intensity between the images We use only first order images derivatives, which are computed using sub-pixel Gaussian filtering.
At section 3, we present a generalization of a gray level corner detector to the case of color images. As we use for matching only first order derivatives of images, we need also here precise first order points of interest. The corner detector implemented here is very stable.
At section 4, we are then able to define a new matching scheme of these primitives using the invariants presented at section 2. The goal of this work is to recover the epipolar geometry between two images.
At section 5, we present comparative matching results with the points introduced at section 3, on images which differ from important rotations and from change of intensity. We show through examples that our method does not need important computational cost and gives very good results in regard to the very good precision of the epipolar geometry obtained. It is evident here that the color information greatly improves the matching method, according to the fact that the additional features allow us to use only derivatives till first order.


next up previous
Next: CHARACTERIZING POINTS OF INTEREST Up: Differential Invariants for Previous: Differential Invariants for

Philippe Montesinos
Wed Jun 2 18:06:30 MET DST 1999