Minimal Perspective Autocalibration

Abstract: We introduce a new family of minimal problems for reconstruction from multiple views. Our primary focus is a novel approach to autocalibration, a long-standing problem in computer vision. Traditional approaches to this problem, such as those based on Kruppa's equations or the modulus constraint, rely explicitly on the knowledge of multiple fundamental matrices or a projective reconstruction. In contrast, we consider a novel formulation involving constraints on image points, the unknown depths of 3D points, and a partially specified calibration matrix $K$. For $2$ and $3$ views, we present a comprehensive taxonomy of minimal autocalibration problems obtained by relaxing some of these constraints. These problems are organized into classes according to the number of views and any assumed prior knowledge of $K$. Within each class, we determine problems with the fewest -- or a relatively small number of -- solutions. From this zoo of problems, we devise three practical solvers. Experiments with synthetic and real data and interfacing our solvers with COLMAP demonstrate that we achieve superior accuracy compared to state-of-the-art calibration methods. The code is available at https://github.com/andreadalcin/MinimalPerspectiveAutocalibration

• The paper introduces a new autocalibration method that solves for unknown focal lengths using minimal correspondence problems.
• It leverages polynomial equations and Gröbner basis solvers, improving reliability in two-view and multi-view reconstructions.
• The approach enhances 3D reconstruction accuracy, benefiting applications such as robotics, augmented reality, and computer vision.

Understanding the Paper: Camera Autocalibration Revisited

Overview

In the world of computer vision, autocalibration of cameras is crucial for achieving accurate 3D reconstructions from 2D images. The paper "Camera Autocalibration Revisited" by A.P. Dal Cin and T. Pajdla tackles this important problem by revisiting some classical methods and proposing new approaches for special cases of camera calibration, focusing on scenarios involving unknown focal lengths.

Background Concepts

Before diving into the findings, let's clarify some underlying concepts:

• Camera Calibration: The process of determining the internal camera parameters (focal length, principal point, etc.) and external parameters (position and orientation of the camera).
• Projective Depth: The scale factor that projects the 3D coordinates to 2D image plane coordinates.
• Fundamental Matrix ($\mathbf{F}$): Encodes the epipolar geometry of a pair of views.
• Essential Matrix ($\mathbf{E}$): Derives from the fundamental matrix but assumes the cameras are calibrated.

The paper addresses scenarios where not all camera parameters are known, particularly focusing on cases where the focal length is unknown.

Two-view 6-point Problem (2V6P)

Problem Setup

Consider two cameras where we know their relative positions and orientations but have an unknown common focal length $f$, and we observe six corresponding points ($N=6$). The challenge is to estimate the relative pose and the common focal length simultaneously.

Key Equations

• Epipolar Constraint: $\mathbf{x}^T \mathbf{F} \mathbf{x}' = 0$
• Connection to Essential Matrix: $\mathbf{E} = \mathbf{K}^T \mathbf{F} \mathbf{K}$, where $\mathbf{K}$ is the camera calibration matrix.
• Demazure Equations for $\mathbf{E}$: $\mathbf{E}$ has rank 2 and two equal non-zero singular values.

For this setup, the fundamental matrix is parameterized and several polynomial equations in unknowns $x, y$, and $w = 1/f^2$ are derived, commonly solved using Gröbner basis solvers.

Three-view 4-point Problem (3V4P)

Euclidean Reconstruction

In a scenario with three cameras and four points, the aim is Euclidean reconstruction. Given three images of four points, we derive

$2NP = 24 \geq 3P + 6N - 7 = 23$

The polynomials are expressed involving the projective depths and inter-point distances. The paper discusses symmetric and asymmetric dehomogenization techniques for solving this system.

Generalized Problems

3 View 5 Point (3V5P) and 6 Point (3V6P)

For these problems, the paper discusses scenarios where there are multiple unknown intrinsic parameters. The system of equations grows more complex, and balancing equations and unknowns becomes challenging. Methods of selective polynomial omission and minimal problem setups are illustrated for robust calibration.

Implications of the Research

This paper's findings offer new methodologies and revisions of older techniques for camera autocalibration, optimizing the calibration process under various scenarios:

• Special Case Handling: Enhancements over traditional methods, especially for cases with unknown focal lengths.
• Reliability in Practical Scenarios: Provides frameworks to ensure that solutions to over-constrained polynomial systems are bounded and manageable.
• Applicability: Broadly applicable to tasks in 3D modeling, augmented reality, and robotics where precise camera calibration is critical.

Future Directions

Here are some potential future directions stemming from this work:

1. Scalability: Extending these methods for calibrating multiple cameras in larger setups.
2. Real-Time Applications: Exploring efficient real-time algorithms for dynamic scene reconstruction.
3. Robustness to Noise: Investigating the resilience of these methods against real-world noise and imperfections in data.

Conclusion

In summary, "Camera Autocalibration Revisited" offers valuable insights and advances in the field of camera autocalibration, with a strong focus on problem setups involving unknown focal lengths. The methods presented could play a pivotal role in improving the accuracy and efficiency of 3D reconstruction algorithms in various domains.

By leveraging these advanced calibration techniques, practitioners can achieve more consistent and precise reconstructions, fostering advancements in numerous computer vision applications.