OpenMVG-rs

A Rust-native computer vision toolkit for Multiple View Geometry and Structure from Motion (SfM): from photos of a scene taken at different viewpoints, it recovers where each camera stood and a 3D point cloud of the scene. Use it from Python, from the command line, or as a Rust library; results export to PLY and COLMAP.

• Incremental and Global SfM: Incremental adds images one at a time; Global solves all their poses at once.
• Deterministic: the same images and seed give the same reconstruction at every thread count.
• Memory safe: every library crate forbids unsafe code, and the default build has no native dependency.
• Open source: dual-licensed under MIT and Apache-2.0.

Quick start

Make a new folder and download a small example dataset, 11 photographs of Sceaux Castle:

mkdir sceaux-demo
cd sceaux-demo
git clone https://github.com/openMVG/ImageDataset_SceauxCastle

From Python, make this folder a uv project and add the package:

uv init --bare --python 3.12
uv add openmvg-rs

Save this script as sceaux.py in the same folder:

import openmvg_rs

scene = openmvg_rs.reconstruct("ImageDataset_SceauxCastle/images")
print(scene)  # <openmvg_rs.Reconstruction engine='incremental' registered=11/11 points=...>
scene.write_ply("sceaux.ply")

Then run it:

uv run sceaux.py

With pip instead, inside a virtual environment (Python 3.11 or newer): python -m pip install openmvg-rs, then python sceaux.py. Open sceaux.ply in a point-cloud viewer such as MeshLab or CloudCompare. The Python guide covers reading the results as NumPy arrays, COLMAP export, the other options, and getting good results.

From the command line, the openmvg-cli package installs the openmvg-sfm command:

cargo install openmvg-cli
# writes ImageDataset_SceauxCastle/sfm_output/reconstruction.ply
openmvg-sfm ImageDataset_SceauxCastle

openmvg-sfm --help lists the engines, camera models, and outputs, and cargo install openmvg-cli --features tui adds openmvg-sfm-configure, which groups the images into cameras interactively; see the openmvg-cli README.

From Rust, use the library through the umbrella crate, which re-exports every layer:

[dependencies]
openmvg = { version = "0.1", features = ["sfm-image-io", "io"] }

Reconstruct a folder of images, then save the points and cameras as a PLY point cloud and as a COLMAP model:

use openmvg::io::write_colmap;
use openmvg::sfm::{Dataset, Incremental};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let dataset = Dataset::from_dir("ImageDataset_SceauxCastle/images")?;
    let output = Incremental::default().run(&dataset)?;
    let reconstruction = output.reconstruction();
    println!("{} views registered", reconstruction.registered_view_count());
    reconstruction.write_ply("scene.ply")?;
    write_colmap(reconstruction, "colmap")?;
    Ok(())
}

Crates

Crate	Description	Release posture
openmvg	Umbrella crate re-exporting core, geometry, features, sfm, and io	Recommended entry point
openmvg-core	Camera models, IDs, graphs, poses, typed units, and deterministic seeds	Stable foundation
openmvg-geometry	F/E/H, pose, alignment, triangulation, averaging, and robust estimation	Stable numerical library
openmvg-features	Detection, description, matching, and tracking	Stable feature library
openmvg-sfm	Dataset, prepared ViewGraph, Incremental and Global reconstruction, and reports	Stable workflow library (umbrella feature sfm)
openmvg-ba	Checked bundle adjustment (BA) and its nonlinear numerical primitives	Stable numerical library
openmvg-io	Reader, writer, and validator of the OMVG community format (manifest plus Parquet tables)	Draft format (umbrella feature io)
openmvg-cli	The openmvg-sfm command, and the openmvg-sfm-configure terminal app behind tui	Command-line programs (source install)

The optional Rerun visualization adapter, openmvg-rerun, is built separately; see DEVELOPER.md. The reproducibility contract defines what identical scientific output means and how RNG protocol changes are accepted.

Learn

Each crate's documentation opens with what the crate is for, a runnable example, and a learning path through its modules; the algorithm pages explain each method with a diagram, its failure modes and its references. Read them in pipeline order:

1. openmvg: which module serves which goal, and how the crates fit together.
2. openmvg-core: cameras, poses, and the conventions every crate shares.
3. openmvg-features: keypoints, descriptors, matches and tracks.
4. openmvg-geometry: from matches to poses and 3D points.
5. openmvg-sfm: the Incremental and Global engines.
6. openmvg-ba and openmvg-io: bundle adjustment, and storing and sharing results.

To browse the documentation locally:

pixi run cargo doc --no-deps --features openmvg/full --open

Building from source

Development uses Pixi to provide the pinned Rust toolchain and native dependencies.

# Clone the repository
git clone https://github.com/openMVG/openmvg-rs
cd openmvg-rs

# Build all crates
pixi run cargo build --workspace

# Run tests
pixi run cargo test --workspace

# Check documentation
pixi run cargo doc --workspace --no-deps

Python

The openmvg-rs package (import openmvg_rs) runs the same reconstruction from Python. Its README lists the platforms with wheels, how to build it from source, and the whole API.

Documentation

The documentation index links to crate guides, the OMVG community format, architecture, validation and contributor instructions.

AI and coding agents

Parts of this project were developed with assistance from large language models, especially for complex math, optimization, and testing.

Citation

If you use OpenMVG-rs in a publication, please cite it together with the original OpenMVG paper. CITATION.cff holds the same metadata, and GitHub turns it into a "Cite this repository" entry.

@software{openmvg_rs,
  title   = {{OpenMVG-rs}: Multiple View Geometry and Structure from Motion in Rust},
  author  = {Moulon, Pierre and {OpenMVG-rs contributors}},
  year    = {2026},
  version = {0.1.0},
  url     = {https://github.com/openMVG/openmvg-rs}
}

@inproceedings{moulon2016openmvg,
  title     = {{OpenMVG}: Open Multiple View Geometry},
  author    = {Moulon, Pierre and Monasse, Pascal and Perrot, Romuald and Marlet, Renaud},
  booktitle = {International Workshop on Reproducible Research in Pattern Recognition (RRPR)},
  pages     = {60--74},
  year      = {2016}
}

Please also cite the papers behind the algorithms you rely on; each solver's module documentation lists its provenance.

License

Licensed under either of the Apache License, Version 2.0 or the MIT license at your option.

Contributing

Contributions are welcome. Read the contributing guidelines before opening a pull request.