Homaloidal parametrization for detecting critical two-view configurations¶
Conference: CVPR 2026
Paper: CVF Open Access
Code: https://github.com/rakshith95/degeneracy-homoloidal
Area: 3D Vision / Multi-view Geometry
Keywords: Fundamental Matrix, Critical Configurations, Degeneracy Detection, Quadratic Transformation, Homaloidal Net
TL;DR¶
This paper proposes a novel quadratic transformation parametrization for detecting general critical surface degeneracies in two-view configurations using the "homaloidal net of conics" from projective geometry. By solving a linear system with 7 image correspondences to fit the quadratic transformation and using an 8th point for verification, the method identifies degeneracies without pre-estimating the fundamental matrix. It achieves higher precision and approximately 200× faster performance than the only comparable method by Luong–Faugeras.
Background & Motivation¶
Background: Estimating the fundamental matrix \(F\) from image correspondences is a core step in Structure-from-Motion (SfM) pipelines. However, when the scene geometry falls into "critical configurations," the estimation of \(F\) becomes unstable or theoretically non-unique. Typical degeneracies include pure camera rotation, all 3D points being coplanar with the camera centers, or all 3D points and camera centers lying on a ruled quadric (e.g., hyperboloid of one sheet, cone, cylinder), known as a "critical surface."
Limitations of Prior Work: In practice, only the "planar scene" case has a widely adopted degeneracy detector based on the homography matrix \(H\) (since \(H\) is a linear mapping between image points and easy to estimate). For general critical surfaces, a practical detector has been missing: the only existing general method is Luong–Faugeras [19], which estimates an initial \(F_P\), then solves a non-linear iterative optimization to find a second \(F_Q\), forming a quadratic transformation for verification.
Key Challenge: Luong–Faugeras has two fundamental issues: (1) It depends on a pre-estimated \(F_P\), which is ill-conditioned near critical configurations, creating a circular dilemma of needing to estimate the most unreliable quantity to detect the degeneracy. (2) Non-linear iterative optimization is slow and often fails to converge to the global optimum.
Goal: To design a stable and practical degeneracy detector for general critical surfaces that avoids pre-estimating the fundamental matrix, relies only on linear systems, and can determine degeneracy directly from 8 correspondences.
Key Insight: A known result in geometry states that 3D points on a critical surface are related across two images by a quadratic transformation \(\Phi\) (a plane-to-plane birational mapping with 14 degrees of freedom). The problem thus becomes how to fit this \(\Phi\) stably and linearly.
Core Idea: Parametrize \(\Phi\) using the "homaloidal net of conics" from classical algebraic geometry—a tool never previously used for degeneracy detection. Through this parametrization, fitting \(\Phi\) reduces to solving a linear system, bypassing fundamental matrix estimation entirely.
Method¶
Overall Architecture¶
The input consists of 8 point correspondences \(\{(x_i,y_i)\}_{i=1}^{8}\) (in homogeneous coordinates), and the output is a binary decision: whether these correspondences originate from a critical configuration. The pipeline involves randomly selecting 7 points to solve a \(14\times 7\) linear system for the quadratic transformation \(\Phi\) using homaloidal net parametrization. The 8th point is then projected via \(\Phi\) to measure the reprojection error \(\mathrm{err}(x_8,y_8)=\lVert y_8-\Phi(x_8)\rVert\). If the error is below a threshold \(\epsilon\), the configuration is deemed degenerate; otherwise, it is non-degenerate and safe for 8-point \(F\) estimation. No fundamental matrix is computed during this process.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input: 8 Point Correspondences<br/>{(xi, yi)}"] --> B["Homaloidal Net Parametrization<br/>Reduce Φ from 18 parameters to 14 DoF"]
B --> C["Linear Fitting of Φ<br/>Solve 14×7 linear system using 7 points"]
C --> D["Degeneracy Test<br/>Error of 8th point ‖y8 − Φ(x8)‖"]
D -->|Error < ε| E["Decision: Degenerate (Critical Configuration)"]
D -->|Error ≥ ε| F["Decision: Non-degenerate<br/>Proceed with 8-point F estimation"]
Key Designs¶
1. Homaloidal Net Parametrization: Explicitly extracting the 14 DoF of \(\Phi\)
A general quadratic plane-to-plane mapping \(\Phi\) consists of three quadratic homogeneous polynomials in \(x_0,x_1,x_2\) (Eq. 1), totaling 18 parameters. Linear fitting would require at least 9 points—more than estimation of \(F\). However, as a birational quadratic transformation, \(\Phi\) actually has only 14 degrees of freedom. This paper uses the algebraic structure of homaloidal nets to extract these 14 DoF.
Specifically, a quadratic transformation can be represented by a pair of bilinear equations \(y_0L_0+y_1L_1+y_2L_2=0\) and \(y_0M_0+y_1M_1+y_2M_2=0\) (Lemma 1), where \(L_i, M_i\) are linear forms of \(x\). The key observation (Prop. 3) is that the homaloidal net remains invariant under left-multiplication of the \(2\times3\) coefficient matrix by any \(2\times2\) invertible matrix \(H\). Consequently, four parameters (\(l_{00},l_{10},m_{00},m_{10}\)) can be fixed to random values, leaving exactly 14 free parameters. This algebraic insight turns the abstract DoF count of \(\Phi\) into a concrete, solvable parameter set.
2. 7-point Linear Fitting of \(\Phi\): A \(14\times 7\) linear system without \(F\)
With the 14-parameter model, each correspondence \((x_i,y_i)\) contributes two linear constraints from the bilinear equations (Eq. 12). Seven points yield 14 constraints, forming a linear system \(Az=b\) (Algorithm 1). The design matrix \(A\) is constructed from monomials like \([x_0y_2,x_1y_0,\dots,x_2y_2]\), and \(b\) contains constant terms derived from the fixed parameters. The system is exactly determined for 7 points and can be solved using least squares for more points.
This is the essential difference from Luong–Faugeras, which first estimates up to three \(F_P\) matrices using the 7-point algorithm and then applies lsqnonlin for each to fit \(F_Q\) (geometric transfer error in Eq. 4). The proposed method is more stable because it avoids the ill-conditioned \(F\) estimation in critical regions.
3. 8th-point Degeneracy Test: Converting degeneracy to a reprojection error threshold
Since 7 points can always fit \(\Phi\) regardless of criticality, the discriminative power comes from the 8th point. The error is defined as \(\mathrm{err}(x_8,y_8)=\lVert y_8-\Phi(x_8)\rVert\) (Eq. 13). If all 8 points lie on a critical surface, they obey \(\Phi\) with minimal error; otherwise, the error is significantly larger. The threshold \(\epsilon\) is comparable to image noise and similar to RANSAC thresholds for \(F\). The authors highlight that F1 scores remain near 1.0 across a wide range of thresholds, indicating stability.
Loss & Training¶
This method is a geometric solver and requires no training. It is implemented in MATLAB and tested on a laptop with an i5-14500HX CPU and 16GB RAM.
Key Experimental Results¶
Experiments were primarily conducted on synthetic critical configurations (consistent with [6, 11, 19]), supplemented by real images. Critical surfaces were modeled using hyperboloids of one sheet, where points were projected using randomly generated camera pairs. Luong–Faugeras [19] served as the primary baseline.
Main Results: Zero-noise Stability + Classification F1¶
| Scene | Metric | Ours | Luong–Faugeras [19] | Note |
|---|---|---|---|---|
| 8th-point Error (100 trials) | Median Error | 9×10⁻¹⁵ | 3×10⁻⁵ | Ours: [1e-16, 7e-12], LF: [6e-15, 0.96] |
| No-noise Classification | F1 | 1.0 (Wide range) | Always < 1 | LF fails due to unstable \(F\) estimation |
| Runtime (Single test) | Time | 8.5 ms | 1.65 s | Approx. 200× Speedup |
Ablation Study¶
| Configuration | Key Observation | Note |
|---|---|---|
| 3D distance from critical surface \(\theta\in[10^{-20},10^0]\) | Our error rises smoothly with \(\theta\); LF error is large/high variance even at small offsets | \(F\) is ill-conditioned near critical regions, affecting LF |
| 2D correspondence Gaussian noise \(\sigma\in[10^{-20},10^0]\) | Both degrade with noise, but Ours maintains lower error and F1≈1 over a wider range | Verification of robustness |
| Real Image (Cylindrical plaza layout) | 8th-point error ~4×10⁻⁵ pixels; correctly classified as degenerate | Validation on real-world critical geometry |
| Real Planar Scene | Coplanar error err≈1.43×10⁻⁵; General position err≈0.94 | Universal for degeneracies where quadrics degrade to planes |
Key Findings¶
- Stability stems from bypassing \(F\): All instabilities in LF arise from pre-estimating \(F\) at critical points. Ours solves directly from image correspondences, leading to a median error roughly 10 orders of magnitude lower than LF.
- Threshold \(\epsilon\) Insensitivity: In zero-noise conditions, F1=1 covers a broad threshold range, meaning precise tuning is not required for deployment.
- Universality: While homography detection is specific to planes, the quadratic transformation test holds for any critical surface, including those that degenerate into planes.
Highlights & Insights¶
- Applying Algebraic Geometry to Vision: The homaloidal net of conics is a novel addition to degeneracy detection. Using its invariance (Prop. 3) to reduce \(\Phi\) from 18 to 14 parameters turns a non-linear problem into a linear one—a strategy applicable to other birational mapping problems.
- Minimalist "Leave-one-out" Criterion: 7-point fitting + 8th-point error check compresses complex degeneracy detection into a simple reprojection threshold, making it easy to implement (millisecond speed, 200x faster).
- Paradigm for Ill-conditioned Problems: Instead of estimating an ill-conditioned quantity, the authors substitute it with an equivalent but well-conditioned mapping (\(\Phi\) instead of \(F\)).
Limitations & Future Work¶
- The experiments rely heavily on synthetic data; the real-world validation is qualitative and lacks a large-scale quantitative benchmark.
- The method is designed for non-minimal cases (≥8 points); the 7-point minimal case (which has a different definition of degeneracy) is out of scope.
- The output is a binary decision; it does not yet extract specific geometric parameters (like surface type or plane+parallax) for reconstruction. Future work may explore the behavior of \(\Phi\) when points are coplanar and its utility in 3D reconstruction.
- ⚠️ OCR artifacts: The quadratic transformation \(\Phi\) and some indices in Eq. 1 and Eq. 12 may have minor typesetting discrepancies in the source; refer to the original paper for precise coefficient arrangements.
Related Work & Insights¶
- vs Luong–Faugeras [19]: Same task and assumptions. LF estimates \(F_P\) then fits \(F_Q\); Ours uses homaloidal net to solve \(\Phi\) linearly without \(F\). Ours is more stable (10 orders of magnitude lower error), faster (200x), and threshold-insensitive.
- vs Homography / Model Selection (GRIC [27], DEGENSAC [8,14]): Those methods focus on pure rotation or planar degeneracies (using linear \(H\)). Ours covers general critical surfaces (using quadratic \(\Phi\)), providing a more universal criterion.
- vs Minimal 7-point Case [11,12]: Minimal cases define degenerate loci on Riemannian manifolds and require curve tests or homotopy continuation for implicit functions. Ours is a practical, linearized approach for non-minimal cases.
- vs Machine Learning / Deep Classification [3,22,23]: Those treat degeneracy as a black-box classification problem. Ours is purely geometric, requiring no training or 3D reconstruction.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ First introduction of homaloidal nets to degeneracy detection.
- Experimental Thoroughness: ⭐⭐⭐ Strong synthetic analysis; needs larger real-world quantitative datasets.
- Writing Quality: ⭐⭐⭐⭐ Clear geometric derivation, though requires some background in algebraic geometry.
- Value: ⭐⭐⭐⭐ Provides a stable, practical, and millisecond-level degeneracy detector for SfM.