tetra3rs

A fast, robust lost-in-space star plate solver written in Rust.

Given a set of star centroids extracted from a camera image, tetra3rs identifies the stars against a catalog and returns the camera's pointing direction as a quaternion — no prior attitude estimate required. The goal is to make this fast and robust enough for use in embedded systems such as star trackers on satellites.

Documentation: For tutorials, concept guides, and Python API reference, see the tetra3rs documentation. For Rust API docs, see docs.rs.

Features

• Lost-in-space solving — determines attitude from star patterns with no initial guess
• Tracking mode — when an attitude hint is available (e.g. the previous frame's solution), skip the 4-star pattern-hash phase and match centroids directly against catalog stars near the hinted boresight. Succeeds with as few as 3 stars, robust to sparse/low-SNR fields, with automatic fallback to lost-in-space if the hint is stale
• Fast — geometric hashing of 4-star patterns with breadth-first (brightest-first) search
• Robust — statistical verification via a per-star likelihood ratio with a bounded false-positive probability
• Multiscale — supports a range of field-of-view scales in a single database
• Proper motion — propagates catalog star positions to any observation epoch
• Compact binary databases — databases serialize with postcard in a portable, lightweight format with no offset-size limit, so wide-FOV-range multiscale databases of any size load cleanly
• Centroid extraction — detect stars from in-memory pixel data with local background subtraction, connected-component labeling, and quadratic sub-pixel peak refinement. Accepts an already-decoded image::DynamicImage or a raw &[f32] pixel buffer (requires the image feature)
• Camera model — unified intrinsics struct (focal length, optical center, parity, distortion) used throughout the pipeline
• Distortion calibration — fit either a SIP polynomial or a standard camera-intrinsics model (free optical center, focal scale, and full Brown-Conrady k1, k2, k3, p1, p2 — the same parameter set as OpenCV's calibrateCamera) from one or more solved images via calibrate_camera, with alternating per-image WCS refinement and global sigma-clipped LS fit. Handles mosaic cameras (e.g. TESS) where the optical axis sits far off the detector
• WCS output — solve results include FITS-standard WCS fields (CD matrix, CRVAL) and pixel↔sky coordinate conversion methods
• Stellar aberration — optional correction for the ~20" apparent shift in star positions caused by the observer's barycentric velocity, with a built-in convenience function for Earth's barycentric velocity

Installation

Rust

The crate is published on crates.io as tetra3:

cargo add tetra3

Python

Binary wheels are available on PyPI for Linux (x86_64, ARM64), macOS (ARM64), and Windows (x86_64):

pip install tetra3rs

To build from source (requires a Rust toolchain):

pip install .

Quick start

Star catalog

tetra3rs uses a merged Gaia DR3 + Hipparcos catalog. The merged catalog uses Gaia for most stars and fills in the brightest stars (G < 4) from Hipparcos where Gaia saturates.

Python: The catalog is bundled in the gaia-catalog package (installed automatically with tetra3rs). No manual download needed — just call generate_from_gaia() with no arguments.

Rust: Download the pre-built binary catalog (G < 10, ~17 MB):

mkdir -p data
curl -o data/gaia_merged.bin "https://storage.googleapis.com/tetra3rs-testvecs/gaia_merged.bin"

Or generate your own with a custom magnitude limit. Two scripts live in scripts/:

• download_gaia_catalog.py — ESA Gaia Archive TAP server. Canonical Gaia source, but the server enforces a hard 3,000,000-row output limit per async job for anonymous users, so bulk queries top out at G ≈ 11.5.
• download_gaia_flatiron.py — Flatiron Institute's flathub service. Serves the full Gaia DR3 catalog without the faint-end cap; use this if you need G > 11.5. flathub is not on PyPI — install it from the upstream repo.

Both produce byte-compatible output and share the Hipparcos bright-star merge. See the Star Catalog docs page for setup, CLI flags, and the on-disk format.

Example

use tetra3::{GenerateDatabaseConfig, SolverDatabase, SolveConfig, Centroid};

// Generate a database from the Gaia catalog
let config = GenerateDatabaseConfig {
    max_fov_deg: 20.0,
    epoch_proper_motion_year: Some(2025.0),
    ..Default::default()
};
let db = SolverDatabase::generate_from_gaia("data/gaia_merged.bin", &config)?;

// Save the database to disk for fast loading later
db.save_to_file("data/my_database.bin")?;

// ... or load a previously saved database
let db = SolverDatabase::load_from_file("data/my_database.bin")?;

// Solve from image centroids (pixel coordinates, origin at image center)
let centroids = vec![
    Centroid { x: 100.0, y: 200.0, mass: Some(50.0), cov: None },
    Centroid { x: -50.0, y: -10.0, mass: Some(45.0), cov: None },
    // ...
];

let solve_config = SolveConfig {
    fov_max_error_rad: Some((2.0_f32).to_radians()),
    // 15° horizontal FOV estimate, 1024×1024 image
    ..SolveConfig::new((15.0_f32).to_radians(), 1024, 1024)
};

// The solve returns Result<Solution, SolveFailure>; a Solution's
// fields are all guaranteed present.
if let Ok(solution) = db.solve_from_centroids(&centroids, &solve_config) {
    println!("Attitude: {}", solution.qicrs2cam);
    println!("Matched {} stars in {:.1} ms",
        solution.num_matches, solution.solve_time_ms);
}

Algorithm overview

1. Pattern generation — select combinations of 4 bright centroids; compute 6 pairwise angular separations and normalize into 5 edge ratios (a geometric invariant)
2. Hash lookup — quantize the edge ratios into a key and probe a precomputed hash table for matching catalog patterns
3. Attitude estimation — solve Wahba's problem via SVD to find the rotation from catalog (ICRS) to camera frame
4. Verification — project nearby catalog stars into the camera frame, match them to centroids, and accept only if the likelihood ratio bounds the false-positive probability below threshold
5. Refinement — re-estimate the rotation using all matched star pairs via iterative SVD passes
6. WCS fit — constrained 3-DOF tangent-plane refinement (rotation angle θ + CRVAL offset) with sigma-clipping, producing FITS-standard WCS output (CD matrix, CRVAL)

Parity flip detection

Some imaging systems produce mirror-reflected images (e.g. FITS files with CDELT1 < 0, or optics with an odd number of reflections). In these cases the initial rotation estimate yields a reflection (determinant < 0) rather than a proper rotation. The solver detects this by checking the determinant of the rotation matrix; when negative, it negates the x-coordinates of all centroid vectors and recomputes the rotation.

The SolveResult includes a parity_flip flag (bool / True/False in Python) indicating whether this correction was applied. This is critical for pixel↔sky coordinate conversions: when parity_flip is True, the mapping between pixel x-coordinates and camera-frame x must include a sign flip.

Tracking mode

When you already have a rough attitude estimate — typically the previous frame's solution in a video-rate star tracker, a propagated gyro estimate, or a coarse attitude sensor — you can skip the lost-in-space pattern-hash phase entirely by passing an attitude_hint:

Rust:

use tetra3::SolveConfig;

// Reuse the camera model from the previous solve so the refined focal length carries over.
let config = SolveConfig {
    attitude_hint: Some(prev_solution.qicrs2cam),
    hint_uncertainty_rad: 1.0_f32.to_radians(),
    ..SolveConfig::with_camera_model(prev_solution.camera_model.clone())
};
let result = db.solve_from_centroids(&centroids, &config);

Python:

# `attitude_hint` accepts either a 4-element [w, x, y, z] quaternion
# (Hamilton, scalar-first) or a 3×3 rotation matrix.
result = db.solve_from_centroids(
    centroids,
    fov_estimate_deg=15.0,
    image_shape=(1024, 1024),
    camera_model=prev_result.camera_model,
    attitude_hint=prev_result.quaternion,  # or .rotation_matrix_icrs_to_camera
    hint_uncertainty_deg=1.0,
)

The solver projects catalog stars near the hinted boresight, nearest-neighbor matches them to centroids, and runs the same Wahba SVD + verification + WCS refine path as lost-in-space. Tracking succeeds with as few as 3 matched stars (lost-in-space needs 4) and is robust to pattern-hash failures from sparse / low-SNR fields. On failure it falls back to lost-in-space automatically — set strict_hint=True (strict_hint: true in Rust) to opt out of the fallback.

See docs/concepts/tracking.md for details on hint uncertainty, quaternion convention, and limitations.

Stellar aberration correction

Stellar aberration is the apparent displacement of star positions caused by the finite speed of light combined with the observer's velocity — analogous to how rain appears to fall at an angle when you're moving. For Earth-based observers, this shifts apparent star positions by up to ~20" (v/c ≈ 10⁻⁴ rad). Without correction, the solved attitude is biased by up to ~20".

To correct for aberration, pass the observer's barycentric velocity (ICRS, km/s) via SolveConfig::observer_velocity_km_s. The solver applies a first-order correction (s' = s + β − s(s·β)) to all catalog star vectors before matching and refinement, producing an unbiased attitude.

The convenience function earth_barycentric_velocity() provides an approximate Earth velocity using a circular-orbit model (~0.5 km/s accuracy, sufficient for the ~20" effect):

Rust:

use tetra3::{earth_barycentric_velocity, SolveConfig};

// days since J2000.0 (2000 Jan 1 12:00 TT)
let v = earth_barycentric_velocity(9321.0);

let config = SolveConfig {
    observer_velocity_km_s: Some(v),
    ..SolveConfig::new((10.0_f32).to_radians(), 1024, 1024)
};

Python:

from datetime import datetime
import tetra3rs

v = tetra3rs.earth_barycentric_velocity(datetime(2025, 7, 10))
result = db.solve_from_centroids(
    centroids,
    fov_estimate_deg=10.0,
    image_shape=img.shape,
    observer_velocity_km_s=v,
)

Catalog support

Tests

Unit tests run with the default feature set:

cargo test

Integration tests require the image feature and test data files. Test data is automatically downloaded from Google Cloud Storage on first run and cached in data/:

cargo test --features image

SkyView integration test

Solves 10 synthetic star field images (10° FOV) generated from NASA's SkyView virtual observatory, which composites archival survey data into FITS images at any sky position. These use simple CDELT WCS (orthogonal, uniform pixel scale). Each image is solved and the resulting RA/Dec/Roll is compared against the FITS header WCS.

cargo test --test skyview_solve_test --features image -- --nocapture

TESS integration test

Solves Full Frame Images (~12° FOV) from NASA's TESS (Transiting Exoplanet Survey Satellite), a space telescope that images large swaths of sky to detect exoplanets via stellar transits. TESS images have significant optical distortion and use CD-matrix WCS with SIP polynomial corrections. The science region is trimmed from the raw 2136×2078 frame to 2048×2048 before centroid extraction.

The test suite includes:

• 3-image basic solve — solves each image and verifies the boresight is within 30' of the FITS WCS solution.
• 3-image distortion fit — fits a 4th-order polynomial distortion model from each solved image, re-solves, and verifies the center pixel RA/Dec is within 1' of the FITS WCS solution.
• 10-image multi-image calibration — solves 10 images from the same CCD (Camera 1, CCD 1) across different sectors with 4 tiered solve+calibrate passes (progressively tighter match radius and higher polynomial order). After calibration, all 10 images achieve RMSE < 9" and center pixel agreement with FITS WCS < 3".

cargo test --test tess_solve_test --features image -- --nocapture

Credits

This project is based upon the tetra3 / cedar-solve algorithms.

• cedar-solve — Steven Rosenthal's Python plate solver, which this implementation closely follows for the star quad generation and matching. (excellent work!)
• tetra3 — the original Python implementation by Gustav Pettersson at ESA
• Paper: G. Pettersson, "Tetra3: a fast and robust star identification algorithm," ESA GNC Conference, 2023

License

MIT License. See LICENSE for details.

This project is a derivative of tetra3 and cedar-solve, both licensed under Apache 2.0 (which in turn derive from Tetra by brownj4, MIT licensed). The upstream license notices are included in the LICENSE file.