| name | calibration-benchmark-and-ablation |
| description | Benchmark and ablation playbook for comparing calibration methods fairly. Use this when asked to compare baselines, evaluate a new candidate, design experiments, or decide whether an algorithm improvement is real. |
Calibration Benchmark and Ablation
Use this skill to compare calibration methods fairly and decide what actually improved.
Core policy
Benchmarking must answer:
- does the candidate beat the baseline?
- where does it beat the baseline?
- where does it lose?
- is the gain robust enough to matter operationally?
Do not benchmark only on a friendly sequence or a single scalar metric.
Benchmark design
Always define:
- Method set
- trusted baseline, candidate, and optionally a stronger oracle-like upper bound
- Dataset matrix
- easy, nominal, and failure-prone cases
- if possible, both in-repo data and a public or standardized reference set
- Perturbation plan
- seed offsets, weak-data cases, or synthetic perturbations
- Metric set
- accuracy, robustness, repeatability, runtime, diagnostic completeness
- Decision rule
- what counts as a real win
Required experiment types
Prefer to include:
- nominal comparison
- holdout comparison
- perturbation recovery
- seed sensitivity
- data thinning or sample-reduction stress
- runtime or operator-cost comparison
Ablation policy
When a candidate has multiple changes, isolate them:
- screening change alone
- initialization change alone
- objective change alone
- refinement change alone
This prevents accidental credit assignment to the wrong idea.
Required outputs
When writing a benchmark plan or result summary, include:
- Methods compared
- Data matrix
- Metrics table
- Win-loss summary
- Notable failure cases
- Recommendation
Good comparison habits
Prefer:
- the same artifact contract for all methods
- fixed acceptance metrics across methods
- explicit "no result" accounting when a method fails
- per-case tables, not only mean scores
- visual review on representative wins and losses
LiDAR-to-LiDAR benchmark matrix
For lidar2lidar, include at least these case families when possible:
- high-overlap direct pairwise scenes
- loop-capable rigs with healthy perimeter edges
- no-loop rigs with several static windows
- wall-dominant or low-corner scenes
- seed-perturbed reruns
- static publish-skew cases vs truly motion-rich timing cases
Report whether the candidate:
- improves pairwise quality
- improves loop consistency when loops exist
- reduces or increases solution-family instability
- keeps the same diagnostics and artifact contract as the baseline
Reference canon
Good anchors include:
- Open3D registration evaluation examples
- PCL registration examples
- Kalibr for report-style quality review
- public datasets and community benchmarks where relevant, but only when the sensor setup is meaningfully comparable
- engineering practice from production autonomy stacks where robustness beats flashy average metrics
Anti-patterns
Avoid these benchmarking mistakes:
- averaging away catastrophic failures
- evaluating on only one recording
- changing metrics between methods
- silently dropping failed cases
- claiming SOTA from internal-only, unreproducible wins
Output style
Lead with the recommendation:
- keep baseline
- promote candidate behind a flag
- replace baseline
- inconclusive
Then justify it with win-loss structure, not just a mean metric.