Abstract
This paper presents the first explicit calculation of the Chow form for the essential variety, a central object in 3D computer vision. By utilizing Ulrich sheaves and representation theory, the authors derive a $20 \times 20$ skew-symmetric matrix formula. This formula provides a mathematical certificate for the physical consistency of six point correspondences between two images, offering a robust alternative to heuristic consistency checks in Structure from Motion problems.
Introduction
In computer vision, the "Structure from Motion" (SfM) problem involves reconstructing a 3D scene from a sequence of 2D images. A fundamental component of this process is the essential matrix, which encodes the relative geometry between two calibrated cameras. The set of all possible essential matrices forms the essential variety.
While the "5-point algorithm" is the standard method for estimating camera geometry from five point pairs, using six pairs creates an overdetermined system. A critical question arises: Given six point correspondences, how can we determine if they are physically consistent with a real 3D configuration? Mathematically, this consistency is captured by the "Chow form"—a single high-degree polynomial that vanishes if and only if the points are consistent. Prior to this work, the Chow form for the essential variety was known to exist but was considered too complex for explicit representation or computation.
Variety Mapping and Geometric Lifting
The authors address this complexity by employing a "lifting" technique from algebraic geometry. They demonstrate that the essential variety can be viewed as a specific hyperplane section (a "slice") of a simpler variety: the set of $4 \times 4$ symmetric matrices of rank at most 2, denoted as $X_{4,2}^s$.
As shown in Table 1, these varieties have well-defined geometric properties. By relating the essential variety to $X_{4,2}^s$, the authors can leverage the structured nature of symmetric matrices to simplify the derivation of the Chow form. The variety $X_{4,2}^s$ has degree 10 and codimension 3, providing a more manageable starting point than the essential variety directly.
Methodology: Ulrich Sheaves and Representation Theory
To represent the geometry of the variety in a computationally tractable way, the authors construct two "Ulrich sheaves." These are special vector bundles that allow a variety to be described through linear algebra—specifically as the determinant or Pfaffian of a matrix.
The construction involves several advanced steps:
- Representation Theory: Using Schur modules and Pieri’s rule, the authors analyze the symmetry of the maps within the sheaf's resolution. This analysis proves that the resulting matrix is skew-symmetric.
- Eisenbud-Schreyer Construction: This method is applied to convert the abstract sheaf into an explicit $20 \times 20$ matrix $M(x, y)$, where the entries are linear forms in the coordinates of the image points.
This approach transforms a "giant" polynomial of degree 120 in 24 variables into a compact matrix representation. If written out term-by-term, the polynomial would be impossible to store or evaluate efficiently; the matrix acts as a highly compressed version.
Results and Numerical Stability
The primary result is Theorem 1.1, which provides the explicit $20 \times 20$ skew-symmetric matrix $M(x, y)$. A set of six points is consistent if and only if this matrix is rank-deficient. Because the matrix is skew-symmetric, the Chow form is represented as its Pfaffian.
To test the practical utility of this formula, the authors performed simulations using synthetic data generated via the Macaulay2 software package. They corrupted exact 6-point correspondences with varying levels of uniform noise (from $10^{-1}$ to $10^{-15}$). The results demonstrated significant numerical robustness: the ratio of the 18th to the 20th singular value ($\sigma_{18}/\sigma_{20}$) scaled linearly with the noise level. This "spectral gap" provides a reliable, noise-aware mechanism for detecting outliers in real-world datasets.
Conclusion and Future Work
This work provides a new determinantal description of the essential variety, simplifying its algebraic study and offering a practical tool for consistency checking. While the $20 \times 20$ matrix is compact by algebraic standards, its integration into high-speed RANSAC loops remains a challenge for future optimization. Additionally, the success of the Ulrich sheaf construction in this context suggests it may be applicable to other complex varieties encountered in computer vision and geometric modeling.
Summary updated