An improvement of LQS (Least Quantile of Squares) algorithm, with an image bucketing to choose samples of data located in the whole image, is used: the idea is to estimate the fundamental matrix with samples of data (here the matched corners), using enough samples to be sure that the probability is high (95 to 99 %) that at least one sample contains no false match; the fundamental matrix that fits best the whole data according to a robust criterion, here the quantile (an improved median), is used. Finally, an estimation is performed on data from which elements considered to be outliers by the previous statistical analysis have been deleted (RLS ou Reweighted Least Squares) [Strom93].

For each sample, as well as during the RLS step, a parameter estimation is to be performed.

Three criteria are used:

• the (well conditionned) 8-point algorithm;
• the virtual parallax method;
• the minimisation of the distance to the epipolar line.

• the matrix elements; 
• Couapel parameters [Couap94];
• Luong parameters [LDFP93].