Robust Geometric Fitting via Outlier Rejection

This direction is split into other directions: outlier sensitivity (outlier sensitivity in geometric modeling). Its last version stays readable and citable here.

Version 1 of 129 papers29 conceptsfrozen 2026-10-04

In computer vision and robotics, "geometric estimation" is the process of fitting a mathematical model—such as a camera’s pose, a 3D shape, or a sensor's coordinate frame—to observed data points. While the math behind these models is well-understood, real-world data is rarely "clean." It is often contaminated by outliers: incorrect data points caused by sensor noise, signal interference, or faulty feature matching.

Standard estimation techniques, like the method of least squares, are notoriously fragile. Because they attempt to minimize the squared distance of all points to the model, a single extreme outlier can "pull" the entire model away from the true solution. To build reliable systems for autonomous drones or 3D reconstruction, researchers are developing frameworks that can automatically identify and ignore these erroneous points during the optimization process.

The Geometry of Sparsity

One of the most powerful ways to think about outlier rejection is through the lens of sparsity. In a typical dataset, we assume that most points are "inliers" (valid data) and only a few are outliers. If we treat the problem of geometric fitting as a selection task—deciding which points belong to the model and which do not—we can use optimization techniques designed to find sparse solutions.

Xingyu Jiang and Jiayi Ma proposed a framework called Dual Sparsity Pursuit (DSP). Instead of just estimating parameters, DSP treats geometric fitting as a sparse subspace recovery problem. By simultaneously identifying the model type and rejecting outliers, the system becomes robust even when the majority of the data is noise. This "pursuit" of a sparse set of valid correspondences allows the algorithm to "see through" the clutter of false matches.

Reformulating the Loss Function

If the standard squared-error loss is too sensitive to outliers, the logical solution is to change the loss function itself. Researchers are moving toward "threshold-resilient" estimators that stop penalizing a point once its error exceeds a certain distance.

Tianyu Huang and Laurent Kneip explore Global Truncated Loss Minimization, which effectively "truncates" the influence of points that are too far from the estimated model. Similarly, in the context of deep learning, Zheng Dang and Pascal Fua have developed eigendecomposition-free training. By avoiding unstable numerical operations and using losses that directly compare predicted matrices against ground truth, these networks can maintain high accuracy in rotation and translation estimation even when most input correspondences are outliers.

Leveraging Geometric Constraints

Another strategy for robustness is to bake the physical constraints of the problem directly into the algorithm. For example, when localizing a network of mobile sensors, we know the sensors must exist in a 3D space governed by distance.

Keyou You and Pei Xie demonstrated that complex, non-convex coordinate alignment can be solved by projecting data onto spherical surfaces. This geometric approach is not only faster but provides a clearer mathematical path for identifying "multipath" interference—outliers caused by signals bouncing off walls—which would otherwise skew the sensor's perceived location.

In the realm of 3D reconstruction, Linfei Pan and Daniel Barath use circular regression to handle rotation averaging. By aligning camera orientations with gravity, they reduce the complexity of the search space, making it easier for the system to distinguish between consistent camera rotations and noisy, inconsistent measurements.

Multi-Model Fitting and Quantum Frontiers

The challenge intensifies when a scene contains multiple geometric models—for instance, detecting several different moving objects in a video. Traditional methods often struggle if they don't know the number of models in advance.

To solve this, Saurabh Pandey and Luca Magri introduced R-QuMF, which reformulates multi-model fitting as a "maximum set coverage" problem. By translating this into a format compatible with quantum annealers, they can search for the best combination of models that explains the most data points while explicitly discarding outliers as "uncovered" noise.

Why It Works: The Intuition of Consensus

Across all these papers, the underlying intuition is consensus. An outlier is, by definition, a point that does not agree with the geometric "story" told by the majority of the data. Whether through sparse optimization, truncated losses, or quantum set coverage, these methods work by finding the largest subset of data that can be explained by a single, consistent mathematical model. By prioritizing this consensus over a "perfect" fit of every single point, these algorithms achieve the stability required for real-world deployment.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_robust_geometric_fitting_via_outlier_rej_2698b6d7,
  title = {Robust Geometric Fitting via Outlier Rejection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-geometric-fitting-via-outlier-rejection/rdr_2698b6d7cbb8478a8ef283486521560b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_geometric_fitting_via_outlier_rej_2698b6d7_v1,
  title = {Robust Geometric Fitting via Outlier Rejection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-geometric-fitting-via-outlier-rejection/rdr_2698b6d7cbb8478a8ef283486521560b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}