In this paper, we have defined a new method for matching stereoscopic uncalibrated color images.
First, we have presented two new differential invariants against Euclidean transformation, specific
to color information.
We have shown that the
addition of the first order Hilbert's invariants computed for each color channel and these
two new invariants gives
eight Euclidean color invariants of first order.
We have then shown that we are able to normalize locally the images to obtain also
affine change of image intensity invariance.
We have shown that considering only these eight first order invariants
is a a sufficient information
to get get a good characterization of the image signal.
In this way color information gives a very simple and stable way to
characterize points of interest of an image.
Considering this characterization we have designed a color detector of points
of interest based on a generalization of the Harris corner detector.
This detector also
needs only first order derivatives.
The results obtained show that this detector is
stable.
We have defined a complete stereoscopic color images matching scheme using the points of
interest and the characterization obtained.
This method is able to match robustely and efficiently high number of points
extracted from color images.
To validate all this approach we show various results on the
Euclidean and affine invariance of our characterization.
We show results on color corner extraction.
Then we show results on the
computation of
the epipolar geometry of a system of two color images differing from
viewpoint and intensity. The accuracy of the recovered geometry shows the validity and the pertinence
of our approach.