Abstract
Classical Cayley-Chow forms provide a polynomial encoding for projective varieties by describing their intersections with linear spaces. This paper extends this theory to varieties within products of projective spaces, establishing the necessary dimension inequalities for unique identification. A significant contribution is the unification of this algebraic framework with "multifocal tensors" in computer vision, explaining the mathematical constraints that limit these tensors to four-camera systems.
Introduction
In algebraic geometry, a variety $X$ in a single projective space can be uniquely represented by its Chow form—a single multihomogeneous polynomial. While this tool is foundational for intersection theory, many modern applications, particularly in computer vision, involve varieties embedded in products of projective spaces, such as $X \subseteq \mathbb{P}^{n_1} \times \dots \times \mathbb{P}^{n_k}$. For instance, a 3D point observed by multiple cameras defines a "multiview variety." This research provides a unified theory for multigraded Cayley-Chow forms, bridging the gap between abstract algebraic geometry and practical multi-view geometry.
Methodology: Incidence and Dimension
The authors define the multigraded Cayley-Chow form through an incidence correspondence. This involves identifying the set of $k$-tuples of linear spaces $(L_1, \dots, L_k)$ whose product $L_1 \times \dots \times L_k$ intersects the variety $X$.
To ensure the Chow form exists as a hypersurface and uniquely determines $X$, the variety must satisfy specific dimension inequalities. The authors utilize polymatroid theory to manage the combinatorial complexity of these requirements. The support of a variety's multidegree is linked to submodular functions, providing a geometric way to visualize the constraints on the slicing planes.
Furthermore, the authors prove that the coordinate rings of products of Grassmannians are Unique Factorization Domains (UFDs). This algebraic property is crucial as it guarantees that the generalized Chow form is indeed defined by a single multihomogeneous polynomial.
Key Insights: The Multifocal Connection
One of the paper's most impactful revelations is that the multifocal tensors used in computer vision—specifically the fundamental matrix, trifocal tensor, and quadrifocal tensor—are precisely the multigraded Cayley-Chow forms of the multiview variety.
This connection explains a long-standing observation in computer vision: multifocal tensors exist for up to four cameras but do not extend to five or more. The paper demonstrates that for $k > 4$, the dimension inequalities required for the Cayley-Chow form to uniquely determine the variety are violated. This provides a rigorous mathematical explanation for the "hard limit" at $k=4$.
Key Results
- Theorem 1.1: Establishes the necessary and sufficient conditions for the set of intersecting planes to form a hypersurface that uniquely identifies $X$.
- Multidegree Formula: Derives a concrete formula for the multidegree of the Cayley-Chow form based on the multidegree coefficients of the original variety.
- Characteristic 0 Stability: Confirms that in characteristic 0, the map from the incidence correspondence to the Chow variety has a multiplicity of 1, consistent with classical theory. Interestingly, in positive characteristics, this multiplicity can exceed 1, marking a departure from the single-projective-space case.
Conclusion and Future Directions
While the multigraded Chow form successfully encodes $X$, the reconstruction process may yield "spurious" components—additional points not present in the original variety. Future work may explore "generalized cameras" (e.g., those capturing lines), where the resulting tensors would manifest as higher-degree forms. The constraint $k \leq r+1$ remains a fundamental boundary for this approach, defining the limits of variety recovery through linear slicing in product spaces.
Summary updated