NemoLeiYANG / MRO

Monocular Rotational Odometry with Incremental Rotation Averaging and Loop Closure

MRO

Author: Chee-Kheng Chng

Given monocular images, MRO estimates their absolute orientations. It is

1. robust against pure rotation motion (equipped with the relative rotation estimator from openGV (https://laurentkneip.github.io/opengv/))
2. capable of loop closing
3. a constant time rotation avergaing solver (due to its incremental nature).

It is an extension to IRotAvg from https://github.com/ajparra/iRotAvg.

License

MRO is released under a GPLv3 license.

For a closed-source version of MRO (e.g., for commercial purposes), please contact the author.

For an academic use of MRO, please cite C.-K. Chng, Á. Parra, T.-J. Chin, Y. Latif: Monocular Rotational Odometry with Incremental Rotation Averaging and Loop Closure, DICTA 2020

Dependencies

Tested on:

• Eigen 3.3.4 sudo apt install libeigen3-dev

• SuiteSparse sudo apt-get install libsuitesparse-dev

• opencv 4.0.0 https://www.pyimagesearch.com/2018/05/28/ubuntu-18-04-how-to-install-opencv/ provides a good guide.

• Boost (Filesystem) 1.6.5 sudo apt-get install libboost-all-dev

• opengv 1.0.0 (refer to the official site for potential installation FAQs)

  • cd third_party/opengv
  • mkdir build
  • cd build
  • cmake ..
  • make
  • cmake -DCMAKE_INSTALL_PREFIX=/path/to/local_install -P cmake_install.cmake

Compilation

Finally, to install MRO

first edit line 8 and 10 in /src/CMakelist.txt to appropriate directories

Then,

• cd /dir/to/MRO
• mkdir build
• cd build
• cmake ..
• make

(binary is compiled inside src)

Demo

The demo.sh bash file in written to run KITTI seq00. The sequence (and the entire dataset) can be downloaded here.

Edit the I/O paths appropriately.

ORB feature extractor used in MRO takes config.yaml and ORBvoc.txt as inputs, both are provided in /data. The current config is tuned according to the camera parameters of KITTI sequence00. It needs to be adjusted accordingly when running on other datasets.

Run ./demo.sh

Hyperparameters

As seen in demo.sh

• MIN_SET : minimal set to run ransac for opengv's relative rotation estimator, original paper proposes 10.

• winsize : the number of neighbouring window of frames that MRO performs feature matching on.

• img_width: image width

• img_height: image height

In /src/IRotAvg.cpp

• rotavg_win_size : number of nodes that the underlying incremental rotation averaging solver takes in during optimisation

• vg_min_matches : minimum number of feature matches to be considered a pair of frames that are viewing the same scene.