Segmenting Epipolar Line

Shengjie Li, Qi Cai, Yuanxin Wu

Abstract

Identifying feature correspondence between two images is a fundamental procedure in three-dimensional computer vision. Usually the feature search space is confined by the epipolar line. Using the cheirality constraint, this paper finds that the feature search space can be restrained to one of two or three segments of the epipolar line that are defined by the epipole and a so-called virtual infinity point.

Introduction

In 3D computer vision, finding "feature correspondences"—identifying the same physical point across multiple images—is a critical yet computationally intensive task. Standard epipolar geometry simplifies this by restricting the search for a matching point in a second image to a 1D line, known as the epipolar line. While this reduces the search from 2D to 1D, searching the entire line remains inefficient and prone to false positives.

Traditional methods like image rectification attempt to simplify this further by warping images so that epipolar lines become horizontal scanlines. However, rectification introduces significant distortion and fails entirely if the epipole (the projection of one camera's center onto the other's image plane) lies within the image frame. This paper proposes a method to narrow the search space on the epipolar line without the destructive effects of rectification.

Methodology

The proposed method leverages the cheirality constraint, a physical requirement that for a point to be visible to a camera, it must have a positive depth (i.e., it must be in front of the lens). By combining this constraint with standard geometry, the authors define a Virtual Infinity Point ($p_{inf}$). This point represents the projection in the second image of a point located at infinity along the viewing ray of the first camera.

Geometric relationship between the epipole (pe), virtual infinity point (pinf), and the epipolar line.

As illustrated in the geometric model above, the epipolar line is naturally divided into segments by the epipole ($p_e$) and the virtual infinity point ($p_{inf}$). By analyzing the rotation and translation components of the camera pose, the authors derive logical propositions to identify which specific segment must contain the true matching point. This effectively treats the epipolar line not just as a mathematical abstraction, but as a representation of a physical ray where light only travels in one direction.

Experimental Results

The method was evaluated using both simulated environments and real-world image pairs. In simulations involving various camera configurations (adjacent, face-to-face, and pure translation), the search space reduction ranged from 31.07% to 66.67%.

Real-world application showing epipolar lines segmented into valid (red) and invalid (blue) regions.

On real-world datasets, the method consistently reduced the search space by approximately 50%. As shown in the figure above, the epipolar lines are segmented into valid regions (red) and invalid regions (blue). Any feature match suggested by algorithms like SIFT that falls on a blue segment can be immediately discarded as a physical impossibility, significantly improving outlier detection and matching speed.

Conclusion

By incorporating the cheirality constraint, this research provides a robust way to segment the epipolar line and halve the search space for feature matching. Unlike rectification, this approach remains effective even when the epipole is within the image frame. While currently requiring calibrated cameras, future work aims to extend this logic to uncalibrated scenarios, further broadening its utility in real-time 3D reconstruction and navigation systems.

Summary updated

Abstract

Identifying feature correspondence between two images is a fundamental procedure in three-dimensional computer vision. Usually the feature search space is confined by the epipolar line. Using the cheirality constraint, this paper finds that the feature search space can be restrained to one of two or three segments of the epipolar line that are defined by the epipole and a so-called virtual infinity point.
BibTeX
@inproceedings{li2020segmenting,
  author = {Shengjie Li and Qi Cai and Yuanxin Wu},
  title = {{Segmenting Epipolar Line}},
  booktitle = {CISP-BMEI},
  year = {2020},
  url = {https://openreview.net/forum?id=sKo15YuH1h},
}
Figure