synthcal 0.1.0

Synthetic dataset generator for camera + laser-stripe calibration.

synthcal

Synthetic dataset generator for camera + laser-stripe calibration.

Install

pip install synthcal
# Optional (recommended for faster Gaussian blur):
pip install "synthcal[scipy]"

Quickstart

Initialize an example config:

synthcal init-config config.yaml

Preview one frame/camera:

synthcal preview config.yaml --frame 0 --cam cam00

Generate a dataset:

synthcal generate config.yaml out_dataset/

python -m synthcal ... is also supported.

Scope

This repository focuses on:

• Generating datasets of static robot poses for an eye-in-hand multi-camera rig.
• For each frame and camera, producing paired outputs from the same pose:
  1. target.png: chessboard under normal illumination
  2. stripe.png: black background with only a laser stripe (when laser is enabled)
  3. corners_px.npy (float32, N x 2) + corners_visible.npy (bool, N)
  4. stripe_centerline_px.npy (float32, M x 2) + stripe_centerline_visible.npy (bool, M)

Units: millimeters (mm) everywhere for geometry. Pixel coordinates are in pixels.

Camera model: OpenCV-style intrinsics K + distortion dist = (k1, k2, k3, p1, p2) (no OpenCV dependency).

Examples

See examples/minimal_no_laser.yaml, examples/minimal_with_laser.yaml, and examples/multicam_medium.yaml.

Camera model

Projection uses the standard pinhole model with distortion applied in normalized coordinates (x, y):

r2 = x^2 + y^2
radial = 1 + k1*r2 + k2*r2^2 + k3*r2^3
x_tan = 2*p1*x*y + p2*(r2 + 2*x^2)
y_tan = p1*(r2 + 2*y^2) + 2*p2*x*y
xd = x*radial + x_tan
yd = y*radial + y_tan

Undistortion is implemented as a simple fixed-point iteration starting from (xu, yu) = (xd, yd) and repeatedly subtracting the forward-model error until convergence (works well for small/moderate distortion and points near the principal point).

Outputs

The dataset layout is described in manifest.yaml (schema v1). Outputs include:

• config.yaml (the normalized config used to generate)
• manifest.yaml (stable schema v1; also includes version: 1)
• per-frame T_base_tcp.npy (identity for all frames unless scenario pose sampling is enabled)
• per-frame/per-camera *_target.png + *_corners_*.npy
• when laser is enabled: per-frame/per-camera *_stripe.png + *_stripe_centerline_*.npy
• placeholder rig/camera YAML files (rig/)

Effects

Rendered images can be post-processed to add simple realism:

• blur_sigma_px: Gaussian defocus blur sigma in pixels
• noise_sigma: zero-mean Gaussian read noise sigma in intensity units (0..255)

Order of operations: blur → noise → clamp → quantize to uint8. Blur uses SciPy (scipy.ndimage.gaussian_filter) when available; otherwise synthcal falls back to a deterministic NumPy implementation (slower).

Determinism: noise is seeded from the global dataset seed and a stable per-output key (frame_index, camera_name, modality) so re-generating with the same config+seed produces identical images. Effects are applied to both *_target.png and *_stripe.png (when present).

Scenario

Optional scenario pose sampling can generate a different T_base_tcp per frame while enforcing in-view constraints (all corners inside the image with a margin). Presets easy|medium|hard provide reasonable default ranges; all randomness is derived from the global seed.

Laser

Laser output is optional. When enabled, synthcal models the laser as a single infinite plane in the TCP frame: n·X + d = 0 (X in mm). The stripe is rendered only where the laser plane intersects the target plane (Z=0 in target frame). If the planes are nearly parallel, synthcal outputs a black stripe image and empty centerline arrays for that (frame, camera).

Reproducibility

• seed is the single global seed recorded in config.yaml and manifest.yaml.
• Noise/effects use deterministic per-output RNG streams derived from (seed, frame_index, camera_name, modality).
• Scenario pose sampling uses a deterministic per-frame stream derived from (seed, frame_index, "__scenario__", "pose").

Public API

Supported library entry points live in src/synthcal/api.py:

from synthcal import generate_dataset, render_frame_preview

CLI

The CLI provides:

• synthcal init-config
• synthcal preview
• synthcal generate

Preview options:

• --no-effects: show the raw render (before blur/noise/quantize)
• --all-cams: show a grid for all cameras
• --show-stripe: with --all-cams, also show a stripe grid when laser is enabled

Coordinate conventions

• Frames (v0): world == base (eye-in-hand only).
• T_tcp_cam maps TCP-frame points into the camera frame.
• T_cam_target maps target-frame points into the camera frame: X_cam = T_cam_target @ [X_target, 1].
• The chessboard target lies in plane Z=0 in the target frame, with outer corner at (0,0,0).
• Inner corners are ordered row-major (rows first, then cols), matching OpenCV’s convention.

Docs

• docs/config_reference.md
• docs/manifest_reference.md

Extending synthcal

Targets, sensors, sampling, and effects are implemented in src/synthcal/. Only the functions in src/synthcal/api.py are intended to be stable for external users.