Sampson Epipolar Error Analysis

• Sampson epipolar error is a first-order, gradient-weighted approximation to the geometric reprojection error, reducing computational complexity in model fitting.
• It is widely applied in multi-view geometry, structure-from-motion, and robust outlier rejection, balancing speed and precision in pose estimation.
• Extensions of the Sampson error incorporate temporal gradients for rolling-shutter cameras, enhancing accuracy under moderate noise conditions.

The Sampson epipolar error is a first-order, gradient-weighted correction to the algebraic epipolar error, providing an efficient approximation to the true geometric reprojection error for two-view correspondences. It is widely used in multi-view geometry, structure-from-motion, and robust matching pipelines as a surrogate cost for model fitting, outlier rejection, and local refinement. The Sampson error balances computational efficiency and geometric fidelity, and recent research has generalized it to encompass more complex camera models, such as rolling-shutter sensors, by augmenting the underlying constraint Jacobian.

1. Classical Sampson Epipolar Error: Formulation and Derivation

For two corresponding homogeneous image points $x_1 = (u_1, v_1, 1)^\top$ and $x_2 = (u_2, v_2, 1)^\top$ and a fundamental matrix $F \in \mathbb{R}^{3 \times 3}$, the algebraic epipolar constraint takes the linear form $x_2^\top F x_1 = 0$. However, realistic observations are noisy and typically violate this constraint. The goal is to quantify the minimal correction needed to project the points back onto the epipolar manifold. The true geometric reprojection error is

$$\min_{\delta x_1,\, \delta x_2} \|\delta x_1\|^2 + \|\delta x_2\|^2, \quad \text{subject to} \ (x_2+\delta x_2)^\top F (x_1+\delta x_1) = 0.$$

Direct minimization is nonlinear and expensive. The Sampson error linearizes this constraint about the current (noisy) observation using first-order Taylor expansion. This gives the explicit closed-form approximation

$$d_S^2(x_1, x_2; F) = \frac{[x_2^\top F x_1]^2} { (F x_1)_1^2 + (F x_1)_2^2 + (F^\top x_2)_1^2 + (F^\top x_2)_2^2 }$$

where $(F x_1)_i$ and $(F^\top x_2)_i$ denote the $i$-th coordinate of the respective 3-vectors, and only the first two dimensions contribute due to the projective geometry structure (Dai et al., 2016, Fathy et al., 2017, Rydell et al., 2024).

The denominator is the sum of squared spatial gradients of the original constraint, encapsulating the sensitivity of the residual to perturbations in both correspondences. This formulation is consistently derived in geometric estimation and substantiated by first-order optimality analysis (Fathy et al., 2017, Rydell et al., 2024).

2. Geometric and Statistical Properties

The Sampson error is a first-order (Gauss–Newton) approximation to the optimal reprojection cost, under the assumption of small measurement noise and low constraint curvature. The approximation holds tightly in the regime where the residual is small or the geometric configuration is regular (Jacobian well-conditioned, Hessian negligible). Theoretical bounds confirm that as the second derivative (constraint curvature) or the residual increases, the bias grows—Sampson underestimates the true error by $O(\varepsilon^3)$ (Fathy et al., 2017, Rydell et al., 2024).

Empirical studies demonstrate that, for typical measurement noise ($\sigma \leq 1$px), the Sampson metric is within 1–2% of the true (optimal) geometric error and retains this tightness up to moderate noise levels (Fathy et al., 2017, Rydell et al., 2024). Sampson also exhibits lower computational complexity, requiring a single matrix-vector product and a few vector norms per correspondence, yielding a $40\times$–$50\times$ speedup compared to nonlinear optimization of the geometric cost (see Table).

Error Metric	Median Error (px)	Relative Speed	Typical Bias
Sampson	0.84	0.35 ms	$\lesssim$2%
Symmetric Epipolar	0.83	0.48 ms	$\lesssim$2%
Full Nonlinear (LM)	0.82	18 ms	—
Pure Algebraic	—	0.12 ms	Large

3. Extensions: Rolling-Shutter Cameras and Generalized Epipolar Geometry

The classical Sampson error assumes global-shutter cameras, but many modern sensors use a rolling-shutter mechanism in which each row exposes at a different time. Dai et al. generalized the essential matrix to a row-dependent "rolling-shutter essential matrix," $E_\mathrm{RS}(y_1, y_2)$, expressible as a polynomial in the row indices $y_1, y_2$:

$$E_\mathrm{RS}(y_1, y_2) = E_0 + y_1 E_1 + y_2 E_2 + y_1 y_2 E_{12}$$

where $E_0 = [t]_\times R$, $E_1, E_2, E_{12}$ encode first- and second-order velocity corrections parameterized by camera scan parameters and instantaneous velocities.

The rolling-shutter Sampson error is

$$e^2_\mathrm{RS}(x_1, x_2, y_1, y_2) = \frac{[ x_2^\top E_\mathrm{RS}(y_1, y_2) x_1 ]^2} { \sum_\mathrm{spatial} + \sum_\mathrm{temporal} }$$

where the denominator sums spatial and newly arising "temporal" gradient terms with respect to row indices. This augmentation penalizes epipolar constraint violation both spatially and temporally, producing a first-order approximation to the true rolling-shutter geometric residual (Dai et al., 2016). If the rolling-shutter parameters (inter-row delay, velocities) vanish, the error reduces identically to the classical global-shutter Sampson form.

4. Applications and Empirical Performance

The Sampson error is the preferred surrogate for two-view model fitting, relative pose estimation, and robust outlier rejection. It is the default cost for RANSAC scoring and also appears in pose refinement and bundle adjustment pipelines when computational resources are limited or real-time feedback is required (Fathy et al., 2017, Rydell et al., 2024).

In practical benchmarks, using the Sampson metric as the inlier criterion leads to robust results nearly indistinguishable from full reprojection error minimization, particularly for moderate to low noise. Applications in multi-camera synchronization (VisualSync) and weakly-supervised correspondence learning (SCENES) leverage the Sampson cost either as a minimization objective (for offset estimation, (Liu et al., 1 Dec 2025)) or as a loss for feature learning and matching (Kloepfer et al., 2024).

Experimental evidence shows that the Sampson metric outperforms simple algebraic or cosine-based alternatives in noise robustness and accuracy, and is competitive or superior to the symmetric epipolar distance in large-scale pose refinement (Liu et al., 1 Dec 2025, Rydell et al., 2024).

5. Relationship to Other Epipolar Error Metrics and Geometric Interpretations

While the Sampson error is a single-term, first-order approximation, alternatives such as the symmetric epipolar distance offer a two-sided, second-order correction and generally more accurate but computationally heavier estimates (Fathy et al., 2017). The symmetric epipolar distance exhibits $O(\varepsilon^4)$ bias, as opposed to $O(\varepsilon^3)$ for Sampson.

In calibrated geometry, the normalized epipolar error, algebraically equivalent to the Sampson numerator, admits several geometric interpretations: (i) it is proportional to the (signed) tetrahedral volume spanned by camera centers and backprojected rays, (ii) to the minimum distance between the two viewing rays, (iii) to the sine of the dihedral angle between epipolar planes, and (iv) to the $L_1$-optimal angular correction required to force ray intersection (Lee et al., 2020). These perspectives clarify when epipolar corrections are geometrically meaningful and when the approximations become ill-posed due to degeneracy (low parallax, near-coplanarity of viewing rays, etc.).

6. Limitations, Assumptions, and Practical Guidelines

The validity of the Sampson approximation is contingent upon small error residuals, well-conditioned geometry, and sufficiently accurate camera calibration. It deteriorates in the presence of large noise, severe constraint curvature, or degenerate configurations (e.g., collinear camera centers or rays aligned with the baseline), where it can significantly underestimate the true geometric error and admit outliers into the inlier set (Fathy et al., 2017, Rydell et al., 2024).

Guidelines derived from accuracy and timing studies include:

• Use the Sampson error for outlier rejection, RANSAC scoring, or real-time systems.
• For final high-precision bundle adjustment, switch to full nonlinear (Levenberg–Marquardt) geometric error minimization, or use the symmetric epipolar distance when seeking improved accuracy with manageable cost.
• Avoid relying on the Sampson error for error magnitudes $\gg 3$px, or where constraint curvature is high.
• In rolling-shutter or non-central camera models, employ the rolling-shutter Sampson extension for improved robustness and accuracy, incurring only moderate computational overhead (Dai et al., 2016).

In summary, the Sampson epipolar error remains an indispensable tool in geometric vision pipelines, offering a theoretically principled and computationally efficient surrogate to true geometric costs, amenable to generalization and robust to a wide range of practical imaging configurations.