In computer vision, the transition from a 3D world to a 2D image typically destroys Euclidean properties like distance and angle. However, certain geometric relationships remain perfectly preserved regardless of the camera's position or lens distortion. The most powerful of these is the cross-ratio, a projective invariant that allows researchers to perform complex tasks—like camera calibration, 3D reconstruction, and geolocalization—without needing prior knowledge of the camera's internal parameters.
The Geometry of Invariance
The core intuition behind this research direction is that while a perspective projection stretches and squashes shapes, it does not do so randomly. If you have four collinear points, their relative spacing changes in the image, but a specific ratio of the distances between them—the cross-ratio—remains identical to the ratio in the physical world.
As Tadahiro Kitahashi and Hiroyuki Endo demonstrated in A New Method of 3-D Motion Analysis Using a Concept of Projective Geometry, this invariance allows for the recovery of 3D spatial relationships from 2D coordinates without frame-by-frame calibration. By identifying features that obey these geometric laws, systems can "undo" the distortion of perspective.
Automating Calibration and Matching
Traditional calibration often requires manual marking of features, but projective invariants enable automation. Guang Jiang and Long Quan utilized the cross-ratio of radii in concentric circles to create a system that "hunts" for circle centers, allowing for high-precision calibration in cluttered scenes without human intervention.
This concept extends to matching features across different views. In Line-sweep: Cross-ratio for wide-baseline matching and 3D reconstruction, Srikumar Ramalingam and colleagues use "virtual lines" between keypoints. By calculating the cross-ratio of points where these lines intersect physical edges, they can verify matches between uncalibrated cameras, effectively bypassing the need for a pre-defined camera model.
Solving for Lens Distortion
Projective geometry is also a potent tool for correcting lens flaws. Because the cross-ratio is a fundamental property of "straightness" in projective space, any deviation from it can be used to measure radial distortion. Y. Liu and Yuchun Huang showed that these invariants can calculate distortion-free coordinates without complex pre-calibration.
More recently, Daniil Sinitsyn and Daniel Cremers introduced PRaDA, which decouples distortion estimation from 3D reconstruction. By operating entirely within projective space, they prove that lens warping is encoded in the relative mapping between image pairs, independent of the scene's actual depth or the camera's focal length.
Real-World Applications: From UAVs to Surgery
The robustness of these invariants makes them ideal for high-stakes environments:
- Geolocalization: Yongfei Li and team use the cross-ratio of road intersections to match UAV imagery to maps in real-time, even when the drone is tilting significantly.
- Medical Imaging: S. Hosseinian developed star-shaped markers for C-arm X-rays that use the golden ratio and projective invariants to maintain accurate AR overlays during surgery.
- Pose Estimation: Juan Carlos Dibene and Enrique Dunn demonstrated that the cross-ratio of "line pencils" (lines meeting at a single point) can recover a camera's absolute pose from just a single scanline of pixels.