Affine reconstruction from a plane and two parallel lines

Fuchao Wu, Guanghui Wang, Zhanyi HuProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2002 · DOI

TL;DR

This paper presents a linear method for affine reconstruction from two images, overcoming a previous impossibility result. By leveraging geometric information like a plane and parallel lines (or parallel planes), the authors enable accurate 3D structure recovery even with a translating camera and changing intrinsic parameters. This work is significant for its practical implications in 3D computer vision.

What Problem Does This Paper Address?

Affine reconstruction, a crucial step in 3D reconstruction, aims to recover the underlying 3D structure of a scene up to a 3D affine transformation. A significant challenge arises when a camera translates and its intrinsic parameters (like focal length or principal point) change between image captures. Previous research showed that without additional geometric information about the scene, it's impossible to achieve affine reconstruction in such scenarios. This paper addresses this limitation by proposing methods that utilize specific scene geometric cues to enable linear affine reconstruction.

Methods

The core of the paper introduces a linear technique for affine reconstruction. The key idea is to exploit geometric constraints present in the scene. Specifically, the authors propose two main scenarios:

  1. Plane and Parallel Lines: If a scene contains a plane and a pair of parallel lines, affine reconstruction can be performed linearly from two images taken by a translating camera. This involves computing the homography induced by the plane and using the vanishing points of the parallel lines to derive necessary parameters.
  2. Parallel Planes (and optionally Parallel Lines): The paper also extends this to cases where parallel planes are present. If only parallel planes are visible, linear reconstruction is possible with a translating camera. If both parallel planes and parallel lines are present, linear reconstruction can be achieved even with more general camera movements (not just translation).

The method relies on calculating the "infinite homography" (induced by the plane at infinity) and the epipole, which are then used to construct the projection matrices for affine reconstruction.

Datasets

  • Simulated Data: The authors generated synthetic data by simulating a pinhole camera model with specified intrinsic parameters and camera movements (pure translation). They introduced varying levels of Gaussian noise (0 to 4 pixels) to test the robustness of their algorithm.
  • Real Images: Experiments were conducted using two images of a calibration rig captured by a digital CCD camera. The rig was set up such that one side served as the "space plane" and two lines on another side represented the "parallel space lines."

Key Insights and Ideas

  • Geometric Priors are Essential: The paper strongly reinforces the idea that for affine reconstruction with a translating camera and varying intrinsics, explicit geometric information about the scene is not just helpful but necessary.
  • Parallel Lines as Vanishing Points: Parallel lines in 3D space project to lines in images that intersect at a vanishing point. This vanishing point provides crucial information about the direction of the parallel lines, which can be leveraged for reconstruction.
  • Planes as Homography Sources: A plane in 3D space induces a homography between its projections in two images. This homography can be computed and used to constrain the reconstruction.
  • Linearity is Achievable: By combining these geometric cues, the authors demonstrate that a linear solution for affine reconstruction is possible, which is computationally efficient and often preferred over non-linear optimization methods.

Key Results

  • Robustness to Noise: Simulations showed that the proposed affine reconstruction algorithm maintains high precision and robustness even under significant Gaussian image noise (up to 4 pixels). The absolute errors and standard deviations for angle (verifying parallelism) and ratio of lengths (verifying line segment preservation) remained low.
  • Preservation of Geometric Invariants: Experiments with real images demonstrated that the reconstructed 3D structure preserved the parallelism of lines and the ratio of lengths on lines, validating the affine reconstruction's accuracy.
  • Generalization: The paper shows that the approach can be extended to handle parallel planes and even general camera views when both parallel planes and parallel lines are present.

Limitations and Future Work

  • Scene Requirements: The primary limitation is the requirement for specific geometric features in the scene: at least one plane and a pair of parallel lines, or parallel planes. This might not always be available in arbitrary scenes.
  • Linearity vs. Euclidean: The reconstruction is affine, meaning it's up to a 3D affine transformation, not a full Euclidean reconstruction (which includes rotation and translation). Recovering the full Euclidean structure would require additional information or assumptions.
  • Future Work: The paper suggests that the proofs for the cases involving parallel planes and general views are omitted, implying these could be further elaborated. Additionally, exploring how to relax the strict geometric requirements or integrate this method with other 3D reconstruction techniques could be future directions.

Summary updated

Abstract

Affine reconstruction is the most difficult and crucial step in 3D reconstruction from image space to the Euclidean space. Recent research indicates that it is impossible to realize the affine reconstruction from a perspective image pairs captured by a translating camera with varying intrinsic parameters if no geometric information of the scene is available. Therefore, it is indispensable to provide some additional information for affine reconstruction from a pair of images. In this paper, we propose that if one plane and a pair of parallel lines are presented in the scene, the affine reconstruction can be done linearly from two images taken by a translating camera. We also point out that if a pair of parallel planes is presented in the scene, the affine reconstruction can also be done linearly from an image pair taken by a translating camera, and if a pair of parallel planes and a pair of parallel lines are presented in the scene, the affine reconstruction can be done linearly from two general views. Extensive simulations and experiments with real images validate our algorithm. The result ofthis paper seems ofboth academic and practical significance in 3D computer vision.
BibTeX
@inproceedings{wu2002affine,
  author = {Fuchao Wu and Guanghui Wang and Zhanyi Hu},
  title = {{Affine reconstruction from a plane and two parallel lines}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2002},
  doi = {10.1117/12.477079},
}
Figure