Filtering Spurious Algebraic Solutions in Geometric Vision

39 concepts26 papers1 proposal

When we model computer vision problems—such as estimating camera pose, calibrating lenses, or reconstructing 3D scenes—we typically translate geometric relationships into systems of polynomial equations. "Minimal solvers" are then used to find exact solutions from a small number of data points.

However, pure algebra is blind to physical reality. Polynomial systems operate over complex numbers and unoriented projective spaces. They do not know that a camera cannot sit inside an object, that focal lengths must be positive, or that an object cannot be transparent. Consequently, solvers frequently yield dozens of mathematically valid but physically impossible "ghost" solutions, as highlighted by Takahashi et al. and Zhang.

To find the true physical state of a scene, geometric vision relies on a suite of physical visibility and plausibility constraints to systematically filter out these spurious algebraic roots.

Chirality and Oriented Projective Geometry

The most powerful tool for eliminating ghost solutions is the concept of chirality (or cheirality)—the requirement that an object must lie in front of the camera lens to be visible.

In standard projective geometry, a line behaves like a continuous loop passing through infinity, which erases the distinction between "front" and "back." As shown by Robert et al., this lack of orientation introduces fundamental ambiguities. For example, Ma et al. demonstrate that failing to account for chirality leaves camera self-calibration equations highly ambiguous, producing multiple mathematically sound but physically inverted results.

  Mathematically Valid (Projective Geometry)       Physically Plausible (Chirality Enforced)

         [Object]                                         [Object]
            |                                                |
            v                                                v
     +--------------+                                 +--------------+
     | Camera Lens  |                                 | Camera Lens  |
     +--------------+                                 +--------------+
            ^                                 
            |                                 
     [Ghost Solution]                         

By enforcing a positive depth constraint, the underlying geometric space can be reduced to a "signed" manifold. This mathematical step naturally resolves mirror-image and orientation ambiguities during multi-view reconstruction, as detailed by Daniilidis et al. and Leonardos.

Post-Filtering vs Early Rejection

How are these physical constraints applied during computation? The strategies generally fall into two categories: post-filtering and early rejection.

Post-Filtering

The traditional approach is to compute all possible mathematical roots first, then discard those that violate physical laws.

  • Trivedi uses physical consistency checks to eliminate non-veridical scene structures from unregistered images.
  • Barath et al. apply visibility checks to reconstructed surface normals, discarding roots that place points behind the camera.
  • Martyushev filters out roots with negative focal lengths or impossible principal points.

Early Rejection

Waiting until the end to filter solutions wastes significant processing power on irrelevant roots, a bottleneck noted by Bujnak et al.. Modern techniques integrate physical constraints directly into the solver or search space to prune hypotheses early:

  • In-Solver Pruning: Persson et al. formulate their Lambda Twist P3P solver to compute point depths directly rather than camera matrices. This allows the algorithm to immediately drop negative depth hypotheses during execution, preventing the refinement of ghost solutions.
  • Search Space Reduction: Li et al. use chirality to segment epipolar lines, cutting the feature-matching search space in half by ignoring segments that represent physically impossible 3D locations behind the lens.
  • Bounding Global Optimizers: In global optimization frameworks, Chandraker et al. use chirality constraints to tightly bound the initial search space of the plane at infinity, accelerating convergence.

Algebraic Saturation: Eliminating Degeneracies at the Ideal Level

Another class of non-physical solutions arises from geometric degeneracies, such as a camera center coinciding exactly with a scene point, or solutions located at infinity. As discussed by Quan et al. and Emiris et al., these spurious solutions inflate the degree of the polynomial system, making the resulting action matrices unnecessarily large and computationally expensive.

To solve this, Larsson et al. introduce the use of saturated ideals from algebraic geometry. By algebraically "dividing out" the ideals corresponding to rank-deficient or infinite camera matrices, the solver completely ignores these uninteresting solutions before computing the action matrix.

As Larsson demonstrates for camera pose estimation with an unknown principal point, this saturation technique eliminates ghost solutions at the foundational algebraic level. The result is a class of minimal solvers that are both orders of magnitude faster and highly stable.

Key Foundational Work

On the Reconstruction of a Scene from Two Unregistered Images (1986) — contributed 2 concept elements to this direction.
Determining Cylindrical Shape from Contour and Shading (1987) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 25 papers
  2. Version 1 · 2026-10-04 · 32 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_physical_visibility_constraints_6f5c067f,
  title = {physical visibility constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physical-visibility-constraints/rdr_6f5c067f49274159905eb345990037a7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_physical_visibility_constraints_6f5c067f_v2,
  title = {physical visibility constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physical-visibility-constraints/rdr_6f5c067f49274159905eb345990037a7/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake