Abstract
Estimating the relative pose between cameras is a cornerstone of computer vision, typically addressed through homographies for planar scenes. While solutions for standard perspective camera pairs are well-established, modern robotics often utilize "generalized cameras"—rigs of multiple cameras with known relative poses. This paper presents the first minimal solvers for the "semi-generalized" case: estimating a homography between one standard perspective camera and one generalized camera. By leveraging five point correspondences and elimination ideal theory, the proposed solvers efficiently recover rotation, translation with absolute scale, and plane parameters, even when the query camera's focal length is unknown.
1. Introduction
In visual localization and navigation, "homography" describes the geometric transformation relating two images of the same planar surface. While central camera models (where all rays pass through a single point) are common, many practical systems—such as autonomous vehicles or pre-mapped environments—are better modeled as generalized cameras. A generalized camera represents a collection of light rays that do not necessarily share a common center.
Registering a new image to such a system is a critical task. When the observed scene is predominantly planar (e.g., building facades or ground planes), homography-based methods are preferred for their robustness. However, the semi-generalized case—one perspective camera versus a generalized rig—has lacked a minimal, efficient solution until now.
2. Problem Formulation and Geometry
The core challenge lies in the asymmetric nature of the setup. As shown in Figure 1, the problem involves a perspective camera $P$ and a generalized camera $G$ (composed of multiple sub-cameras $G_1, G_2, G_3$).
The authors parametrize the homography matrix as $H = R - tn^T$, where $R$ is the rotation, $t$ is the translation, and $n$ is the plane normal. Unlike standard two-view geometry, the internal baseline of the generalized camera allows for the recovery of the absolute scale of the translation $t$, a significant advantage for metric localization.
3. Minimal Solvers via Elimination Ideals
The researchers focus on "minimal" solutions—using the fewest possible data points (five 2D-2D correspondences) to solve the underlying equations. This approach is essential for integration into RANSAC-based robust estimation frameworks.
The technical innovation involves using Elimination Ideal Theory from algebraic geometry. The complex system of polynomial equations derived from the point correspondences is systematically simplified by "eliminating" variables. This process results in a single univariate polynomial of low degree (typically 5 or 3), which can be solved extremely quickly (in microseconds). The authors developed a suite of solvers, including sH52 for fully calibrated rigs and sH5f2 for cases where the query camera's focal length is unknown.
4. Experimental Evaluation
The solvers were rigorously tested against synthetic data and real-world benchmarks, including the KITTI autonomous driving dataset and the Cambridge Landmarks localization dataset.
As illustrated in Figure 4, the proposed solvers exhibit high numerical stability (plots a, e) and maintain accuracy under varying noise levels (plots b, f). A key finding is the solvers' performance in "near-planar" scenes. While standard epipolar geometry often fails when points are close to a plane, these homography solvers remain stable (plots c, g). On the KITTI dataset, a hybrid approach combining these solvers with traditional methods yielded superior rotation and translation estimates.
5. Conclusion
This work bridges a significant gap in camera geometry by providing efficient, minimal solvers for semi-generalized homographies. By enabling absolute scale recovery and handling unknown focal lengths, these solvers provide a robust tool for real-time visual localization in planar environments. Future research may focus on automatically detecting scene planarity to switch between homography and general 3D solvers dynamically.
Summary updated