More Informed RANSAC (MI-RANSAC)

• More-Informed RANSAC (MI-RANSAC) is a refined RANSAC method that pre-tests minimal samples using geometric conditioning to filter ill-posed configurations.
• It employs Riemannian manifold techniques and X.5-point curves to assess sample stability, improving computational efficiency and hypothesis robustness.
• Empirical results demonstrate a 1.5×–2× speed-up in hypothesis generation while maintaining high accuracy in essential and fundamental matrix estimations.

More-Informed RANSAC (MI-RANSAC) is a methodological enhancement to the conventional Random Sample Consensus (RANSAC) framework applied in minimal solver problems for multiview geometry, specifically for the estimation of essential and fundamental matrices in Structure-from-Motion (SfM) pipelines. MI-RANSAC leverages a mathematically principled pre-test to filter out ill-conditioned minimal samples, thereby improving both computational efficiency and hypothesis robustness without measurably compromising accuracy (Fan et al., 2021).

1. Riemannian-Manifold Conditioning and Minimal Problems

The framework situates the minimal problems—namely, the 5-point (essential matrix, E) and 7-point (fundamental matrix, F)—within Riemannian manifolds:

• $\W$: world‐scene manifold
• $\X$: image‐data manifold, being either $(\mathbb{R}^2 \times \mathbb{R}^2)^5$ (5-point) or $(\mathbb{R}^2 \times \mathbb{R}^2)^7$ (7-point)
• $\Y$: epipolar manifold (the set of E or F matrices as appropriate, each a submanifold of $\mathbb{P}(\mathbb{R}^{3\times3})$)

The forward map $\Phi : \W \dashrightarrow \X$ and the epipolar map $\Psi : \W \dashrightarrow \Y$ formalize the generative process. The solution map

$\S = \Psi \circ \Theta : \; \X \supset U \;\to\; \Y, \quad \Theta = \Phi^{-1}$

is defined wherever $D\Phi(w)$ is invertible. The local conditioning of the solution is characterized by the operator norm: $\X$0 where $\X$1 is induced by the Riemannian metrics on $\X$2 and $\X$3.

For the 5-point problem, $\X$4 is a $\X$5 Jacobian; the corresponding condition number is

$\X$6

with $\X$7 the $\X$8 Gram root. In the 7-point variant, an analogous formula involves a $\X$9 Jacobian and a $(\mathbb{R}^2 \times \mathbb{R}^2)^5$0 Gram root $(\mathbb{R}^2 \times \mathbb{R}^2)^5$1 [(Fan et al., 2021), Prop. 3.1–3.2].

2. Geometric Characterization of Ill-Posed Loci

Configurations are classified as ill-posed if $(\mathbb{R}^2 \times \mathbb{R}^2)^5$2 is singular for some corresponding world scene $(\mathbb{R}^2 \times \mathbb{R}^2)^5$3.

• World Scene Description:
  • 5-point: Ill-posed if there exists a rectangular ruled quadric $(\mathbb{R}^2 \times \mathbb{R}^2)^5$4 through the $(\mathbb{R}^2 \times \mathbb{R}^2)^5$5, containing the baseline $(\mathbb{R}^2 \times \mathbb{R}^2)^5$6, and such that all planes orthogonal to $(\mathbb{R}^2 \times \mathbb{R}^2)^5$7 intersect $(\mathbb{R}^2 \times \mathbb{R}^2)^5$8 in circles.
  • 7-point: Ill-posed if a quadric $(\mathbb{R}^2 \times \mathbb{R}^2)^5$9 exists through the seven 3D points, containing the corresponding baseline $(\mathbb{R}^2 \times \mathbb{R}^2)^7$0 [Thms 3.3, 3.4].
• Image Data Description & X.5-Point Curves:
  • 5-point: The last point $(\mathbb{R}^2 \times \mathbb{R}^2)^7$1 lies on a degree-30 real algebraic 4.5-point curve in the image space if and only if the configuration is ill-posed. The defining polynomial $(\mathbb{R}^2 \times \mathbb{R}^2)^7$2 is bidegree 30 in $(\mathbb{R}^2 \times \mathbb{R}^2)^7$3.
  • 7-point: The 7th point $(\mathbb{R}^2 \times \mathbb{R}^2)^7$4 lies on a degree-6 algebraic 6.5-point curve if and only if ill-posed; this can be computed by homotopy or an explicit 1668-term Plücker polynomial [Thms 3.5, 3.6].

The distance from the last correspondence to its respective X.5-point curve, $(\mathbb{R}^2 \times \mathbb{R}^2)^7$5, serves as a robust numerical indicator of proximity to the ill-posed locus.

3. MI-RANSAC Sampling and Pre-Testing Procedure

Rather than sampling minimal subsets indiscriminately, MI-RANSAC incorporates a conditioning pre-test: only subsets sufficiently far from the ill-posed locus are admitted for solver evaluation. This process is summarized for the calibrated (5-point) case as follows:

• Inputs: Set of correspondences $(\mathbb{R}^2 \times \mathbb{R}^2)^7$6, inlier threshold $(\mathbb{R}^2 \times \mathbb{R}^2)^7$7, pre-test distance threshold $(\mathbb{R}^2 \times \mathbb{R}^2)^7$8, max hypotheses $(\mathbb{R}^2 \times \mathbb{R}^2)^7$9.
• For each random 5-subset $\Y$0:
  1. Split $\Y$1 into first $\Y$2 correspondences and the last point.
  2. Compute the 4.5-point curve given the first $\Y$3 points.
  3. Evaluate distance $\Y$4.
  4. If $\Y$5, reject $\Y$6; otherwise, solve the 5-point problem.
  5. For each real root, score the model and retain the hypothesis with maximal inlier count.

The uncalibrated (7-point) procedure analogously uses the 6.5-point curve and a 7th correspondence. Pseudocode for the full process is provided in (Fan et al., 2021), including explicit handling of the curve computation and distance evaluation.

4. Empirical Analysis: Efficacy and Speed-up

Synthetic experiments demonstrate that:

• For 7-point (fundamental matrix), with $\Y$7 px, approximately 90% of ill-conditioned samples are rejected based on the X.5-point distance pre-test, with only ≈10% of stable samples inadvertently rejected.

• For 5-point (essential matrix), a threshold of 2 px similarly distinguishes almost all ill-posed samples from stable ones.

Overall, MI-RANSAC requires roughly half as many minimal-problem solves to generate the same count of well-conditioned hypotheses, yielding computation speed-ups of approximately $\Y$8 to $\Y$9 for the hypothesis-generation stage. The computational load of evaluating the pre-test polynomial ($\mathbb{P}(\mathbb{R}^{3\times3})$0) is negligible compared to the roots-solving step ($\mathbb{P}(\mathbb{R}^{3\times3})$1–$\mathbb{P}(\mathbb{R}^{3\times3})$2) [(Fan et al., 2021), Fig. 9].

5. Practical Implementation Details

Key considerations for integration include:

• Curve Stability: The X.5-point curves exhibit sub-pixel stability ($\mathbb{P}(\mathbb{R}^{3\times3})$3 px movement) when the fixed correspondences are perturbed with realistic noise levels ($\mathbb{P}(\mathbb{R}^{3\times3})$4 px), ensuring pre-test robustness [Fig. 10].
• Polynomial Evaluation: For fundamental matrix estimation, evaluating the 1668-term Plücker polynomial requires at most $\mathbb{P}(\mathbb{R}^{3\times3})$5, with integer coefficients $\mathbb{P}(\mathbb{R}^{3\times3})$6. For the essential matrix, homotopy computation of the high-degree curve is an offline cost; runtime evaluation is reduced to evaluating a univariate degree-30 polynomial.
• Threshold Selection: Empirically, $\mathbb{P}(\mathbb{R}^{3\times3})$7 px (uncalibrated) or $\mathbb{P}(\mathbb{R}^{3\times3})$8 px (calibrated) divides degenerate from stable samples with classification error $\mathbb{P}(\mathbb{R}^{3\times3})$9.
• Alternative Condition Number Estimation: Approximations of $\Phi : \W \dashrightarrow \X$0 via local linear matrix inequalities (LMIs) are feasible, but the X.5-point distance is faster and more direct in practice.
• Integration: MI-RANSAC is orthogonal to other RANSAC accelerations such as PROSAC or guided sampling and can be integrated into any two-view step within an SfM pipeline.

6. Theoretical Context and Implications

The rigorous condition-number formulation draws on Riemannian manifold techniques and aligns with the geometric characterization of minimal solvers’ stability. The investigation of ill-posed loci both in the world scene and in image data, as made explicit by X.5-point algebraic curves, provides a foundational justification for MI-RANSAC’s early rejection mechanism. The distinct separation—quantified in pixels—between stable and unstable configurations underpins the reliability of the pre-test.

A plausible implication is that this methodology could be ported to other minimal-problem domains or serve as a pre-processing layer in non-minimal solvers, provided comparable conditioning criteria can be formulated.

7. Significance and Integration in Structure-from-Motion Pipelines

By precluding near-degenerate samples before minimal solver invocation, MI-RANSAC maintains standard RANSAC’s robustness while delivering a $\Phi : \W \dashrightarrow \X$1–$\Phi : \W \dashrightarrow \X$2 speed-up in hypothesis generation at negligible accuracy loss. This renders it a practical, theory-backed augmentation for SfM and other multiview geometry pipelines reliant on minimal solvers, with immediate empirical and algorithmic impact (Fan et al., 2021).