linear-ransac 🚀

linear-ransac is a Rust implementation of RANSAC (RANdom SAmple Consensus) for fitting a 1D linear model (a line in 2D: y = ax + b), inspired by scikit‑learn’s RANSACRegressor.

It’s built for noisy data with nasty outliers, with automatic thresholding and iteration control so you don’t have to hand‑tune.

Highlights

• Robust line fitting – estimate slope and intercept from 2D points even when a big chunk of them are garbage.
• Auto threshold via MAD – inlier cutoff is computed from the median absolute deviation of y, similar in spirit to scikit‑learn’s default.
• Auto iteration stopping – adapts the max number of RANSAC trials from the current inlier ratio and a stop_probability, while enforcing a configurable minimum so one lucky bad model can’t stop the search early.
• Final OLS refit – once a consensus set is found, refit the line with ordinary least squares on inliers.

Today it targets 1D linear regression; the design is ready to grow to N‑dimensional linear models later on.

Install

[dependencies]
linear-ransac = "0.1"

Quick start

use linear_ransac::{Point, RansacSolver};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let data = vec![
        Point::new(0.0, 0.0),
        Point::new(1.0, 1.0),
        Point::new(2.0, 2.0),
        Point::new(3.0, 3.0),
    ];

    let model = RansacSolver::new().fit(&data)?;

    println!("slope = {} intercept = {}", model.slope, model.intercept);
    Ok(())
}

With outliers:

use linear_ransac::{Point, RansacSolver};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut data = vec![
        Point::new(0.0, 0.0),
        Point::new(1.0, 1.0),
        Point::new(2.0, 2.0),
        Point::new(3.0, 3.0),
        Point::new(4.0, 4.0),
    ];

    // obvious outliers
    data.push(Point::new(1.5, 100.0));
    data.push(Point::new(2.5, -50.0));

    let model = RansacSolver::new().fit(&data)?;

    println!("slope = {} intercept = {}", model.slope, model.intercept);
    Ok(())
}

Tuning

• RansacSolver::new() – defaults tuned for typical use.
• with_seed(u64) – deterministic runs.
• with_min_trials(usize) – guarantee at least this many iterations.

Internally the solver:

• derives an inlier threshold from a plain MAD of y values,
• adapts the iteration budget from the inlier ratio and stop_probability,
• and refits the final line with OLS on the inliers.

Errors

RansacSolver::fit returns a Result<LinearModel, RansacError>:

• InsufficientData – not enough points to define a line,
• ModelFittingFailed – degenerate/numerically unstable data,
• NoConsensusFound – no good consensus set within the trial budget.

License

MIT. See LICENSE.