Anchored VINS Consistency Analysis

This repository contains the companion code to reproduce the figures and tables of our paper: "Observability and Consistency Analysis for Visual-Inertial Navigation with Anchored Feature Parameterizations".

This repo is a fork of OpenVINS and its main purpose is to empirically validate the observability and consistency results derived in the paper. It additionally provides an implementation of the Decoupled Right-Invariant EKF (DRI), based on Yang et al. (2022) - "Decoupled Right-Invariant Error States for Consistent Visual-Inertial Navigation". To use the DRI-EKF, set the parameter consistency_method: "dri_fej" in the config file, which will enable the right-invariant error for the inertial state and pose clones, and utilize the forms of the right-invariant state transition matrix and measurement model Jacobians from Yang et al. (2022).

Installation

Follow the OpenVINS installation guide to install the dependencies. This code was tested with Ubuntu 20.04 and ROS Noetic, using Eigen 3.4.0, OpenCV 4.2.0, and Ceres 2.0.0, and can be built by running

mkdir -p ~/catkin_ws_anchored_vins/src  
cd ~/catkin_ws_anchored_vins/src 
git clone git@github.com:mitchellcohen3/anchored_vins_consistency.git
cd ~/catkin_ws_anchored_vins
catkin build -j4

Reproducing paper results

The scripts to reproduce the figures and tables in the paper are located in the scripts folder, which additionally contains a few scripts that may be useful for analyzing the output of VINS estimators. The Python evaluation scripts use the Python package pyvins, located in the folder python/pyvins. Install the package in a virtual environment using

python3 -m venv .venv
source .venv/bin/activate
pip install -e python/pyvins

which will additionally install all dependencies required to run the evaluation scripts.

To reproduce the results from the simulation experiment, run the script scripts/sim_featrep_experiments/run_camnoise_experiment.sh, which runs Monte-Carlo trials across different camera noise levels for each estimator, on all three datasets. The script will also generate plots of the results in the folder results/plots, used to compare the performance of the different estimators across camera noise levels (Figures I and II from the paper). See also the script scripts/generate_ate_comparison_table.py to generate Table II from the paper.

The script scripts/tumvi_exp/run_tumvi_experiment.sh is used to run the TUM-VI dataset experiments. Download the room sequences ROS bags (in 512x512 resolution) along with the groundtruth files here, and place them in folder with the structure:

dataset_path/
├── dataset-room1_512_16.bag
├── dataset-room2_512_16.bag
...
├── truths
    ├── dataset-room1_512_16.txt
    ├── dataset-room2_512_16.txt
    ....

See the script scripts/tumvi_exp/evaluate_tum_vi_experiment.sh to generate the evaluation plots.

Citation

If you found this repository useful for your research, please cite it as below:

@article{cohen2026observability,
  title={Observability and Consistency Analysis for Visual-Inertial Navigation with Anchored Feature Parameterizations},
  author={Cohen, Mitchell and Korotkine, Vassili and Forbes, James Richard},
  journal={arXiv preprint arXiv:2606.19307},
  year={2026}
}

Acknowledgments and License

This project is built on OpenVINS, released under GPLv3.