Automatic feature matching (using SIFT or ORB) invariably generates outliers—mismatched point pairs that do not follow the underlying geometry. Least-squares estimation is extremely sensitive to these; a single bad match can destroy the entire result.
Random Sample Consensus (RANSAC) solves this by iteratively sampling minimal subsets of data:
The number of iterations $N$ required to find an outlier-free set with probability $p$ is:
Where $e$ is the outlier ratio and $s$ is the sample size.