Synthetic Data Generation

calibration-rs provides utilities for generating synthetic calibration data, used in examples and tests. Synthetic data allows testing with known ground truth, verifying convergence, and debugging algorithm issues. The generators are behind the synthetic feature of vision-calibration (cargo add vision-calibration --features synthetic, or as a dev-dependency feature for tests).

Board Point Generation

#![allow(unused)]
fn main() {
use vision_calibration::synthetic::planar;

// 8×6 grid of 3D points at Z=0, spaced 40mm apart
let board_points = planar::grid_points(8, 6, 0.04);
// Returns Vec<Pt3>: [(0,0,0), (0.04,0,0), ..., (0.28,0.20,0)]
}

Points are generated in the $Z=0$ plane with the specified column count, row count, and spacing (in meters).

Pose Generation

#![allow(unused)]
fn main() {
// 6 camera poses with varying yaw and distance
let poses = planar::poses_yaw_y_z(
    6,      // number of views
    -0.2,   // start yaw (radians)
    0.08,   // yaw step
    0.5,    // start Z distance (meters)
    0.05,   // Z step
);
// Returns Vec<Iso3>: cam_se3_target (board-to-camera) transforms
}

Poses are generated looking at the board center, with rotation around the Y axis (yaw) and varying distance along Z.

Projection

#![allow(unused)]
fn main() {
// Project board points through all poses
let views = planar::project_views_all(&camera, &board_points, &poses)?;
// Returns Vec<CorrespondenceView>: 2D-3D point pairs per view
}

This applies the full camera model (projection + distortion + intrinsics) to generate pixel observations.

Adding Noise

For realistic testing, project with deterministic uniform pixel noise. UniformPixelNoise is a pure function of (seed, view index, point index), so datasets are identical across platforms and library versions:

#![allow(unused)]
fn main() {
use vision_calibration::synthetic::noise::UniformPixelNoise;

let noise = UniformPixelNoise { seed: 42, max_abs_px: 0.5 }; // uniform in [-0.5, 0.5] px per axis
let views = planar::project_views_noisy(&camera, &board_points, &poses, &noise)?;
}

Typical noise levels (max_abs_px):

Set a fixed seed in examples and tests for reproducibility.

Common Test Scenarios

Minimal (3 views, no noise)

For verifying algorithm correctness. Should converge to machine precision.

Moderate (6-10 views, 0.5 px noise)

For testing realistic convergence. Should achieve <1% intrinsics error after optimization.

Challenging (20 views, 1.0 px noise, outliers)

For testing robustness. Should achieve <2% intrinsics error with robust loss functions.

Distortion stress test

Use large distortion ($k_1=-0.3$, $k_2=0.1$) to verify the iterative intrinsics solver handles strong distortion.