Relative Pose Estimation of Calibrated Cameras with Known SE(3) Invariants

A Collection of Algorithms for Relative Pose Estimation of a Calibrated Camera

Introduction

This repository contains the following relative pose estimation solvers, in C++ and Matlab API:

• The conventional 5-point algorithm (5P), wrapped from Hartley's well-known implementation.
• 4-point algorithm with a known rotation angle (4P-RA)
• 4-point algorithm under planar motion without knowing the plane direction (4P-ST0)
• 3-point algorithm with a known rotation angle and under planar motion without knowing the plane direction (3P-RA-ST0).

Compared to other relative pose estimation algorithms, 4P-RA, 4P-ST0 and 3P-RA-ST0 leverage extra sensor/motion constraits without requiring extrinsics calibration. This is due to the interesting property of SE(3) invariants.

Reference

This repository is the source code for the following paper:

@article{li2020relative,
  title={Relative Pose Estimation of Calibrated Cameras with Known SE(3) Invariants},
  author={Li, Bo and Martyushev, Evgeniy and Lee, Gim Hee},
  journal={ECCV},
  year={2020}
}

Contents

• form Source code to derive different solver formulations compared in Table 3 in the paper. Automatic Generator and Generator for Automatic Polynomial Solvers are required to run these code.
• include C++ API.
• matlab Matlab API.
• perf Source code for the experiment section in the paper.

Compilation

The default CMakeLists.txt requires Matlab to be installed to compile the Matlab API.

mkdir build && cd build
PATH=/usr/local/MATLAB/<version>/bin/:$PATH cmake ..
make