THE FOLD / BOSS / THE-RAID / THE RANSAC
THE RANSAC
a model found through a storm of outliers
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
RANSAC (RANdom SAmple Consensus) fits a model to data riddled with outliers — the workhorse of computer vision for finding lines, planes, and geometric relations in noisy point sets. Instead of least-squares (which one bad point can wreck), it repeatedly draws the minimal sample needed to define a model (two points for a line), counts how many other points agree within a tolerance, and keeps the model with the largest consensus set. With enough random trials, it almost surely hits a sample of pure inliers and locks onto the true model — even when nearly half the data is garbage.
LIT verified live: over 800 trials with 60 inliers and 40 outliers, RANSAC recovers the planted line’s slope (to within 0.1) and its inlier set in 100% of runs (window.__ransac). FIG honest scope: RANSAC is randomized — this is a high-probability guarantee, and the success rate is reported, not assumed.
LIT verified live: over 800 trials with 60 inliers and 40 outliers, RANSAC recovers the planted line’s slope (to within 0.1) and its inlier set in 100% of runs (window.__ransac). FIG honest scope: RANSAC is randomized — this is a high-probability guarantee, and the success rate is reported, not assumed.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) seated this at the-raid — the true model survives a raid of outliers by gathering the largest party of points that agree. AVAN (AI) built the instrument: the minimal-sample loop, the consensus count, and the planted-line recovery test.
Credit as content: Fischler & Bolles (1981). The weave: David names the raid; I confirm RANSAC recovers the planted line through heavy outlier contamination, and report the measured success rate rather than overclaiming certainty.
Credit as content: Fischler & Bolles (1981). The weave: David names the raid; I confirm RANSAC recovers the planted line through heavy outlier contamination, and report the measured success rate rather than overclaiming certainty.
3 ONE DIMENSION
Points from a line plus scattered outliers; RANSAC’s best model is the line with the most points in agreement.
4 TWO DIMENSIONS · INTERACTIVE
Generate inliers and outliers; RANSAC finds the consensus line and marks its inliers — least-squares would bend toward the noise.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the recovered line and its consensus set.
AVAN’s addition (the inverse-companion): don’t fit all the points — find the model most points vote for. The inverse of ‘minimize error over everything’ is ‘the line with the largest inlier consensus, outliers ignored.’ Magenta is the outlier storm; green is the line the inliers agree on. Consensus beats least-squares.
LIT Genuine RANSAC (Fischler & Bolles, 1981). Verified live: over 800 trials with 60 inliers (near a planted line) and 40 uniform outliers, the minimal-sample consensus loop recovers the planted slope to within 0.1 and captures ≥90% of inliers in 100% of runs (window.__ransac.successRate).
FIG Honest scope: RANSAC is randomized, so this is a high-probability guarantee — the measured success rate is reported, not assumed certain. The AVAN inverse is honest — instead of minimizing error over all points, one finds the model the most points vote for, ignoring outliers. Magenta is the outlier storm; green is the line the inliers agree on. Consensus beats least-squares.
FIG Honest scope: RANSAC is randomized, so this is a high-probability guarantee — the measured success rate is reported, not assumed certain. The AVAN inverse is honest — instead of minimizing error over all points, one finds the model the most points vote for, ignoring outliers. Magenta is the outlier storm; green is the line the inliers agree on. Consensus beats least-squares.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE-RAID · David Lee Wise (ROOT0), with AVAN