RANSAC fit models by trusting tiny clean samples before noisy data

Classical model fitting is fragile when feature detectors produce many bad points. One wrong cluster can pull a least-squares fit away from the structure that actually exists.

Fischler and Bolles introduced Random Sample Consensus in 1981 as a new paradigm for fitting models to experimental data containing a significant percentage of gross errors.

The paper frames the method for image analysis and automated cartography, including a landmark-location problem: propose a solution from a minimum set of landmarks, then accept the model only if enough other landmarks agree.

Fit the clean minority, then count support

It spends computation on hypotheses, not on trusting every measurement.

Minimal samples make each trial cheap enough to repeat many times.

Consensus converts noisy geometry into a voting problem that outliers struggle to win.

When bad data can dominate averages, sample small and verify broadly. A robust process can be built by making agreement, not residual size, the first filter.

RANSAC became a core computer-vision tool for tasks such as image matching, camera geometry, and localization. OpenAlex lists the 1981 paper with more than 25,000 citations, reflecting how widely the sampling-and-consensus pattern spread.