calibration-capture-design

name	calibration-capture-design
description	Dataset and capture-design guide for sensor calibration pipelines. Use this when asked what data to collect, how to recollect a failing calibration dataset, how to improve observability, or whether a solver problem is really a data problem.

Calibration Capture Design

Use this skill to design calibration data collection before blaming the solver.

When to use

Use this skill when the task involves:

• planning a calibration recording session
• deciding whether poor results require recollection instead of optimizer changes
• improving coverage, excitation, observability, or scene diversity
• defining module-specific capture checklists
• designing a dataset that can support benchmarking and holdout evaluation

Core policy

In calibration work, the highest-leverage fix is often better data rather than a more complex algorithm.

Always design capture around:

1. Required raw signals
   • topics, frames, timestamps, static transforms, and supporting metadata
2. Coverage and excitation
   • viewpoint, depth, tilt, motion direction, turn balance, and scene diversity
3. Observability
   • the data must actually constrain the unknowns of interest
4. Auditability
   • record enough context to explain failures later

Capture design workflow

For any module, specify:

1. what must be recorded
2. what is strongly recommended
3. what motion or scene pattern is needed
4. what weak-data pattern to avoid
5. what fast post-record checks should be run before optimization

Module-specific heuristics

Prefer guidance like:

• camera
  • use the exact production camera mode
  • cover center, corners, and multiple tilts
  • avoid autofocus or auto-exposure drift
• lidar2camera
  • keep the target fully visible in both modalities
  • vary depth, horizontal position, and tilt
  • avoid collecting all poses from one image region or one depth
• lidar2lidar
  • prefer static scenes with walls, corners, poles, and facades
  • reduce moving traffic and open-sky-only segments
  • include geometry that supports repeatable registration
• lidar2imu
  • include both left and right turns
  • include acceleration and braking, not only cruising
  • keep enough flat-road support for ground and motion assessment

Recollection triggers

Recommend recollection before solver redesign when you see:

• low accepted-pair ratio
• weak pose or motion diversity
• poor image or scene coverage
• unstable holdout behavior
• multiple solution families
• geometry that obviously underconstrains the requested degrees of freedom

Required output

When using this skill, include:

1. Capture goal
2. Required signals
3. Planned excitation and coverage
4. Expected observability
5. Fast post-record checks
6. Recollection criteria

Reference canon

Prefer:

• in-repo capture guidance such as docs/apollo_data_collection.md
• OpenCV calibration docs for camera-board coverage practice
• Kalibr workflows for staged collection and validation discipline
• production autonomy workflows that privilege stable data collection over brittle solver heroics

Anti-patterns

Avoid these capture mistakes:

• collecting many redundant frames instead of diverse ones
• recording only nominal straight-line motion for motion-based calibration
• letting auto-exposure or autofocus change during a session
• using scenes with poor geometric structure for targetless registration
• trying to fix an underconstrained dataset with a heavier optimizer

Output style

Lead with whether the current dataset is:

• sufficient
• sufficient for review only
• needs targeted recollection
• fundamentally underconstrained

Then provide the smallest useful recollection plan.