AI Glossary / Computer vision and multimodal AI / Intermediate
Random sample consensus
Definition
Random sample consensus, usually called RANSAC, is a robust model-fitting method that repeatedly fits candidate models to small random subsets and evaluates how many observations agree with each candidate. It aims to estimate a useful model despite the presence of outliers.
What it means
A few incorrect matches can badly distort an ordinary least-squares fit. RANSAC takes a different approach: it looks for a model supported by a sufficiently large consistent subset of the data. Observations close enough to a candidate model are called inliers according to a chosen residual threshold; others are treated as outliers for that candidate. The labels are relative to the model and threshold, not intrinsic proof that an observation is correct or incorrect. RANSAC works best when the intended model has enough support and the sampling procedure can find suitable subsets.
How it works
Choose a model family, a minimum sample size, an error measure and an inlier threshold. Randomly select a minimal or small subset, reject degenerate samples, fit a candidate and evaluate all observations. Keep a candidate according to its consensus or another robust score, and repeat until a stopping rule is met.
A final model is often refitted using the selected inliers. The required number of trials grows when the inlier fraction falls or the sample size increases. The familiar success-probability calculation assumes suitable independent sampling and a fixed inlier fraction. Real implementations may use adaptive stopping, guided sampling or alternative scores, so the exact behavior depends on the variant.
Three practical examples
- A panorama pipeline estimates a homography from feature matches containing incorrect correspondences.
- A point-cloud application fits a plane while rejecting points belonging to other surfaces.
- A line-fitting experiment compares least squares with a consensus-based fit when several observations are far from the intended line.
Think of it this way
Think of asking small groups of witnesses for a candidate explanation, then checking how many other observations support it. Repeating the process can uncover a consistent account despite unreliable observations. The analogy is limited: numerical consistency does not establish truth, and a large coordinated group of wrong matches can support the wrong model.
Worked example
Suppose the inlier fraction is w = 0.5 and fitting the model requires s = 2 sampled observations. Under the usual independent-sampling approximation, a sample is all-inlier with probability w^s = 0.25. The chance that N trials all miss such a sample is (1 − 0.25)^N = 0.75^N.
For at least 99% probability of one all-inlier sample, require 0.75^N ≤ 0.01. This gives N ≥ log(0.01)/log(0.75), approximately 16.01, so 17 trials suffice under those assumptions. An all-inlier sample is not automatically a good final model if it is geometrically degenerate or the noise and model assumptions are unsuitable.
Illustrative case study
In a fictional image-stitching project, repeated windows on a building produce many ambiguous feature matches. RANSAC finds a homography with a large consensus, but the alignment still shifts the pattern by one window. The team inspects spatial coverage and visual correspondence, then improves matching and validation. The example shows that consensus can be misleading when incorrect correspondences are themselves structured. No robustness method should be treated as proof that every surviving match is correct.
Common misconceptions and mistakes
RANSAC does not guarantee the globally best model in a finite random run. A larger threshold can inflate the inlier count by accepting poor matches, while an overly strict threshold can reject valid noisy observations. The method does not automatically choose the correct model family. A planar homography may be unsuitable for a scene with strong depth variation. Another mistake is comparing raw consensus counts without considering spatial distribution, residual quality or competing structures.
Business use
For deployed vision systems, tune thresholds using meaningful residual units and representative noise. Record the randomization and stopping settings when reproducibility matters. Evaluate failure cases with repeated patterns, low inlier fractions, degenerate geometry and multiple valid structures. Review the final residuals and the locations of supporting observations, not just their count. If the estimate controls later processing, provide a clear failure or low-confidence path when the evidence is insufficient.
People also ask
Is RANSAC a neural network?
No. It is a randomized estimation framework and can operate without learned components.
Are all selected inliers correct?
Not necessarily. They satisfy the chosen model's threshold, which can include false matches or exclude valid ones.
Why can it need many iterations?
When all-inlier samples are rare, repeated trials are needed to obtain a suitable candidate with adequate probability.
Check your understanding
Increasing the inlier threshold doubles the consensus count. Does that alone prove the estimated model improved?
Answer: No. The looser threshold may simply admit more incorrect or poorly fitting observations. Residual quality, geometry and task performance still need evaluation.
Related terms
- Outlier
- Residual
- Least squares
- Robust estimation
- inlier
- consensus set
- homography
- reprojection error
- model fitting
- degenerate sample
Learn next
Learn residuals and robust model selection next. Fit a line to a small dataset with controlled outliers, vary the threshold and compare both the model parameters and which observations enter the consensus set.
More in Computer vision and multimodal AI
- Quadrature mirror filter - A quadrature mirror filter, or QMF, is a filter used in a filter-bank arrangement that separates a signal into complementary frequency bands.
- Quantization error - Quantization error is the difference between an original value and the discrete value used to represent it after quantization.
- Radon transform - The Radon transform represents a function through its integrals over lines.
- Ratio test - The ratio test is a feature-matching rule that compares the distance to a descriptor's nearest candidate with the distance to its second-nearest candidate.
- Re-detection - Re-detection is the process of locating a previously tracked target again after tracking has failed, become uncertain or been interrupted.
- Read noise - Read noise is random uncertainty introduced by the electronic process of measuring and reading a sensor's signal.
Sources
OpenCV, official robust homography estimation documentation: https://docs.opencv.org/3.4.5/d9/d0c/group__calib3d.html Carnegie Mellon University, computational-photography lecture discussing RANSAC: https://graphics.cs.cmu.edu/courses/15-463/2017_fall/lectures/lecture20.pdf Primary documentation and teaching-source search excerpts consulted; full source and editorial review remain pending.
Previous: Random rounding Next: Random start