When a 3D scene is projected onto a 2D image plane, metric properties like lengths, angles, and parallel lines are instantly destroyed. A square table becomes an irregular quadrilateral; right angles warp into acute or obtuse angles. This geometric distortion makes tasks like object recognition, feature matching, and camera calibration notoriously difficult when the camera's internal parameters are unknown.
However, projective geometry offers a remarkable loophole: certain algebraic properties remain perfectly preserved under perspective projection. The most fundamental and powerful of these is the cross-ratio. By leveraging cross-ratio invariance, computer vision systems can reason about 3D shapes, match features across wildly disparate views, and calibrate lenses without needing explicit metric information.
Why the Cross-Ratio Works
To understand why the cross-ratio is invariant, we must look at the algebra of perspective projection. If we have four collinear points $A, B, C, D$ along a line, their cross-ratio is defined as a "ratio of ratios" of the distances between them:
$$\text{Cr}(A, B, C, D) = \frac{(C - A)(D - B)}{(C - B)(D - A)}$$
When a camera captures these points, the perspective projection acts as a homography—a linear fractional transformation of the form $f(x) = \frac{ax+b}{cx+d}$. If we compute the distance between two transformed points, the result is:
$$f(x_j) - f(x_i) = \frac{(ad-bc)(x_j-x_i)}{(cx_j+d)(cx_i+d)}$$
When we substitute this back into the cross-ratio formula for four points, the scale factor $(ad-bc)$ and all the camera-dependent denominators $(cx_k+d)$ appear an equal number of times in both the numerator and the denominator. They algebraically cancel out completely.
As early as 1985, Tadahiro Kitahashi et al. demonstrated in A New Method of 3-D Motion Analysis Using a Concept of Projective Geometry that this mathematical rigidity allows algorithms to map 2D image coordinates back to 3D spatial relationships without requiring frame-by-frame camera calibration. Crucially, this invariance applies not just to four points on a line, but dually to a pencil of lines (four concurrent lines intersecting at a single point), where the cross-ratio of the sines of the angles between the lines remains constant.
Robust Feature Matching Across Wide Baselines
In wide-baseline stereo vision, matching features between two cameras placed far apart is incredibly challenging because local visual descriptors warp beyond recognition. Instead of relying purely on visual appearance, researchers use the cross-ratio as a strict geometric validator.
For instance, Srikumar Ramalingam et al. introduced the Line-sweep: Cross-ratio for wide-baseline matching and 3D reconstruction framework. By constructing "virtual lines" between pairs of standard keypoints and finding their intersections with physical edges, they compute cross-ratios that must match exactly across views. This provides a robust verification metric that bypasses the need for prior camera calibration.
View 1 View 2
(Camera A) (Camera B)
\ / \ /
\ / \ /
------- -------
A' B' C' D' A'' B'' C'' D''
\ | / / \ | / /
\|/ / \ | / /
v / \|/ /
|/ v /
| | /
| |/
---------------------------------------------------------
A B C D (3D Collinear Points)
This principle scales up to macro-level problems like drone geolocalization. Yongfei Li et al. developed a method in Road-Network-Based Fast Geolocalization that calculates the cross-ratio of concurrent lines at road intersections. Because these ratios are invariant to the UAV's pitch, roll, and altitude, aerial images can be instantly matched against flat vector maps in under a second.
Even the fundamental matrix—the mathematical bedrock of two-view geometry—can be simplified using this concept. Yoni Kasten et al. showed in Two view constraints on the epipoles from few correspondences that the cross-ratio of four lines within an epipolar pencil remains invariant across different camera perspectives. This allows them to recover epipolar geometry using as few as four to six point correspondences, bypassing the standard seven-point requirement.
Uncalibrated Camera Calibration and Pose Estimation
If the cross-ratio is invariant to camera parameters, it can also be used in reverse: to isolate and solve for those parameters. Traditional calibration requires known 3D coordinates (like a physical checkerboard), but invariants allow self-calibration or pattern recognition without orientation knowledge.
Consider the problem of automatic feature labeling during calibration. Guang Jiang et al. designed a system in Detection of Concentric Circles for Camera Calibration that uses targets composed of concentric circles. Under perspective projection, the centers of these circles shift in the image, but the cross-ratio of their radii remains invariant. This allows the system to uniquely identify and label calibration features from any angle without manual intervention.
Similarly, for medical imaging, S. Hosseinian et al. leveraged cross-ratios within pentagram-shaped fiducial markers in C-ARM POSE ESTIMATION AND NAVIGATION IN SURGERIES FOR AUGMENTED REALITY APPLICATION to achieve fast, viewpoint-independent pose estimation for surgical C-arm X-ray machines.
More recently, Juan Carlos Dibene et al. demonstrated in Camera Resection from Known Line Pencils and a Radially Distorted Scanline that the cross-ratio of a line pencil is preserved even when observed through a single scanline of pixels, allowing for high-frequency absolute pose estimation for rolling-shutter cameras.
Taming Lens Distortion with Projective Geometry
Real-world cameras rarely follow perfect perspective projection due to lens distortion (such as barrel distortion in wide-angle or fisheye lenses). Distortion bends straight lines into curves, seemingly breaking cross-ratio invariance.
However, this "breaking" is exactly how distortion can be measured and eliminated. In CRITICAL ASSESSMENT OF CORRECTION METHODS FOR FISHEYE LENS DISTORTION, Y. Liu et al. utilized the cross-ratio to directly solve for lens distortion coefficients, mapping distorted coordinates back to a space where projective invariants hold true.
A key insight in this domain is that while most lines curve under radial distortion, lines passing directly through the Center of Distortion (CoD) remain perfectly straight. José Henrique Brito et al. exploited this geometric property in Radial Distortion Self-Calibration to locate the CoD across multiple uncalibrated views without any scene assumptions.
Taking this further, Daniil Sinitsyn developed PRaDA: Projective Radial Distortion Averaging, demonstrating that lens distortion can be recovered entirely within projective space, completely decoupled from 3D scene structure or focal length. Because scene geometry is unique up to a homography, the distortion parameters are cleanly encoded within the relative mappings between image pairs.
Beyond Points and Lines: Curves and Surfaces
The power of projective invariants extends beyond simple linear features to complex 3D shapes. David A. Forsyth et al. showed in Recognising rotationally symmetric surfaces from their outlines that when a rotationally symmetric object (like a vase) is projected, the vertices of its bitangent cones form collinear points in the image. The cross-ratio of these projected vertices serves as a viewpoint-invariant signature, allowing for 3D object recognition from a single 2D silhouette.
For non-planar curves, Long Quan established algebraic invariants for pairs of spatial conics in Invariant of a Pair of Non-Coplanar Conies in Space: Definition, Geometric Interpretation and Computation, deriving signatures from matrix pencils that remain constant under any projective transformation.
Even when four collinear points are not naturally visible, virtual geometry can fill the gaps. In Obstacle Detection Using Projective Invariant and Vanishing Lines, Ryuzo Okada combined three physical horizontal lines on a vehicle with a virtual fourth point derived from the road's vanishing line. This created a cross-ratio constraint capable of distinguishing flat road markings from vertical obstacles, proving that projective invariants remain a vital tool for anchoring uncalibrated vision systems.
Key Foundational Work
A New Method of 3-D Motion Analysis Using a Concept of Projective Geometry (1985) — contributed 4 concept elements to this direction.
Recognising rotationally symmetric surfaces from their outlines (1992) — contributed 2 concept elements to this direction.
Projective Invariants for Vision (1992) — contributed 2 concept elements to this direction.