pyvernier 0.0.1.dev1

Vernier — Geometric calibration toolkit for multi-camera and multi-sensor systems

vernier

Geometric calibration, fast. Multi-camera intrinsics, multi-sensor extrinsics, and bundle adjustment through a clean Python API on top of a Ceres Solver / Sophus core.

pip install pyvernier

import numpy as np
import vernier

K = np.array([[1400, 0, 960], [0, 1400, 540], [0, 0, 1]], dtype=np.float64)
distortion = np.zeros(5)

estimator = vernier.Estimator(loss="huber", loss_scale=1.0)
camera = estimator.add_camera("brown_conrady", K, distortion, fixed=["k3"])
frame  = estimator.add_frame(vernier.SE3(), vernier.Submanifold.SE3)

for point_3d, pixel_2d in observations:
    estimator.add_observation(camera, frame, point_3d, pixel_2d)

result = estimator.solve(max_iterations=200)
# SolveResult(converged=True, rms=0.31, iterations=8, time=0.076)

K_opt    = estimator.get_camera_matrix(camera)
dist_opt = estimator.get_dist_coefficients(camera)
pose_opt = estimator.get_frame_pose(frame)

Features

Projection Models

Model	Distortion vector	Use case
BrownConrady	[k1, k2, p1, p2, k3]	Standard rectilinear cameras (OpenCV default)
KannalaBrandt	[k1, k2, k3, k4]	Fisheye / wide-angle
Division	[k1, k2]	Lightweight radial distortion
Rational	[k1..k6, p1, p2]	Complex optics

# String API
camera = estimator.add_camera("kannala_brandt", K, dist)

# Class API
camera = estimator.add_camera(vernier.KannalaBrandt, K, dist)

Per-Parameter Fixing

Hold any subset of intrinsic parameters constant during optimization:

# Fix k3 by name
camera = estimator.add_camera("brown_conrady", K, dist, fixed=["k3"])

# Fix by enum
camera = estimator.add_camera(
    vernier.BrownConrady, K, dist,
    fixed=[vernier.BrownConrady.Param.K3],
)

# Optimize distortion only
camera = estimator.add_camera(
    "brown_conrady", K, dist,
    fixed=["fx", "fy", "cx", "cy"],
)

Submanifold Pose Constraints

Choose which degrees of freedom of an SE(3) frame are optimized:

# Full 6-DOF (default)
frame = estimator.add_frame(pose, vernier.Submanifold.SE3)

# Translation only
frame = estimator.add_frame(pose, vernier.Submanifold.TRANSLATION)

# Rotation only
frame = estimator.add_frame(pose, vernier.Submanifold.ROTATION)

# Arbitrary axis combination
frame = estimator.add_frame(
    pose,
    vernier.Submanifold.X | vernier.Submanifold.Y | vernier.Submanifold.WZ,
)

# Held fixed
frame = estimator.add_frame(pose, vernier.Submanifold.NONE)

Robust Loss Functions

Per-observation loss for outlier handling. Default is Huber (1.0 px) — the standard in bundle adjustment.

Loss	Behavior	When to use
"none"	Pure L2	Clean synthetic data
"huber"	L2 below scale, L1 above	Default — calibration with mild outliers
"cauchy"	Aggressive downweighting	Noisy real-world data
"soft_l1"	Smooth L1 approximation	Dense correspondences

# Estimator-wide default
estimator = vernier.Estimator(loss="cauchy", loss_scale=0.5)

# Per-observation override
estimator.add_observation(camera, frame, p3d, p2d, loss="cauchy", loss_scale=0.5)

Architecture

src/vernier/
├── __init__.py        Public re-exports + version
├── _core.py           Lie groups, projection models, enums, results
└── estimator.py       Estimator — Pythonic problem-construction API

The optimization core wraps Ceres Solver for non-linear least squares with autodiff and Sophus for SE(3) / SO(3) manifold operations. Projection models are stateless C++ trait structs dispatched via std::variant + std::visit so Ceres gets exact Jacobian block sizes per camera.

Documentation

Full documentation will be available at vistralis.org/vernier.

Development

pip install -e ".[dev]"
ruff check .
ruff format --check .
pytest

License

Apache-2.0 — see LICENSE for details.

Copyright (c) 2026 Vistralis Labs.