Abstract: Not available
Introduction
Erwin Kruppa’s seminal 1913 paper, originally titled "Zur Ermittlung eines Objektes aus zwei Perspektiven mit innerer Orientierung," stands as a foundational pillar of geometric computer vision. Predating modern digital imaging by decades, this work provides the first mathematical proof that the relative pose between two calibrated cameras can be determined using only five point correspondences. It introduced what are now known as "Kruppa’s equations," forming the theoretical basis for modern Structure-from-Motion (SfM) and camera self-calibration.
The Problem of Relative Pose
In the early 20th century, photogrammetry was primarily concerned with reconstructing 3D objects from 2D images. At the time, it was known that seven point matches were required to determine the "epipolar geometry" (the geometric relationship between two views) if the cameras were uncalibrated.
Kruppa sought to determine the minimum amount of information needed when the cameras are calibrated—meaning their "inner orientation" or intrinsic parameters like focal length and principal point are known. He aimed to prove that with known calibration, the number of required points drops from seven to five, and to characterize the number of possible 3D solutions that arise from such a minimal set.
Methodology and Geometric Framework
Kruppa utilized the tools of projective and analytical geometry to model the relationship between two views. His approach can be summarized in three main steps:
- Epipolar Line Pencils: He modeled the set of all possible epipolar lines passing through an epipole as a "pencil." For any point in space, its projections in two images must lie on corresponding epipolar lines.
- The Absolute Conic: Kruppa’s key insight was leveraging the "Image of the Absolute Conic" (IAC). In projective geometry, the IAC is a virtual conic at infinity that is invariant to camera displacement. Knowing a camera's calibration is mathematically equivalent to knowing the IAC in that camera's image plane.
- Tangency Constraints: He formulated constraints based on the fact that epipolar lines tangent to the IAC in one image must correspond to epipolar lines tangent to the IAC in the second image. By combining these "conic constraints" with five point correspondences, he derived a system of algebraic equations to solve for the epipoles.
Key Insights and Results
The Five-Point Breakthrough
Kruppa was the first to realize that five points are the "minimal case" for calibrated relative pose estimation. This is a staple of modern computer vision, exemplified by Nistér’s 5-point algorithm. He showed that knowing internal camera parameters provides two additional geometric constraints, allowing researchers to "trade" two point correspondences for calibration data.
Kruppa’s Equations
While named later by Maybank and Faugeras in 1992, the paper outlines the fundamental relationship between the fundamental matrix and the dual image of the absolute conic. This later became the cornerstone of self-calibration, where camera intrinsics are recovered from image sequences without a calibration rig.
Multiplicity of Solutions
Kruppa proved that a 3D configuration of five points can be reconstructed up to scale from two calibrated views. He calculated that there are generally 11 distinct solutions (resulting in 22 possible relative poses, appearing in pairs). Modern refinements (Demazure, 1988) later showed the maximum number of distinct solutions is actually 10. Additionally, he proved that if only one camera is calibrated and the distance between camera centers is known, seven points are sufficient to find up to 24 solutions.
Limitations and Conclusion
While Kruppa proved the existence of these solutions, his method involved solving high-degree polynomials that were computationally impractical before the advent of modern computers. Furthermore, the high number of potential solutions means that in practice, additional points or "cheirality" checks (ensuring points are in front of the camera) are needed to find the single correct physical solution.
Despite these practical hurdles in 1913, Kruppa's work serves as a reminder that many "modern" computer vision problems have deep roots in classical projective geometry. His derivations provided the mathematical scaffolding for the automated 3D reconstruction technologies used today.
Summary updated