Skip to content

Map Cleaning False Deletion Test Protocol

Last updated: 2026-05-09

False deletion is the safety-critical failure mode where a map-cleaning pipeline removes real structure, real hazards, or review-required evidence. This protocol validates that dynamic-object removal does not create an over-cleaned map that looks tidy but is less truthful.

Scope

IncludeExclude
offline point-cloud map cleaning, rejected-layer review, localization replay, semantic map checksruntime obstacle deletion decisions
static assets, movable-static assets, FOD-like objects, unknown/occluded regionsgeneral perception model benchmarking
airside stands, service roads, apron routes, terminal edges, gate equipment zonesfinal airport-wide operating approval by itself

Required Inputs

InputMinimum requirement
Raw survey logsLiDAR packets/frames, camera where available, ego pose, TF, timestamps, calibration
Candidate mapraw accumulated map, cleaned static map, removed layer, restored layer, unknown layer
Ground truthASAM OpenLABEL or equivalent IDs for static, movable-static, dynamic, FOD/hazard, artifact, unknown
Map contexttile ID, route/stand, ODD slice, semantic map version, coordinate frame, release target
Cleaner recordalgorithm, model/config version, thresholds, commit/package hash, operator

Procedure

StepActionPass evidence
1Freeze raw logs, labels, and cleaner configurationimmutable manifest with checksums
2Run cleaner on baseline and holdout sequencesreproducible command and output bundle
3Compare removed layer against ground truthfalse-deletion table by class and zone
4Inspect safety-critical removals manuallyreviewer disposition with screenshot or point-cloud view
5Replay localization on raw and cleaned mapsresidual, covariance, inlier, and relocalization report
6Validate semantic map consistencyLanelet2/vector diff and geometry tolerance report
7Package unresolved riskswaiver, quarantine, rollback, or remediation decision

Test Matrix

ScenarioFalse deletion targetRequired slices
Quiet baseline surveypermanent static assetsdry/day and at least one low-light or wet slice if in ODD
Busy stand surveystatic assets behind aircraft/GSEaircraft present, aircraft absent, GSE staged, GSE removed
FOD placed-object runlow-height hazardssmall metal, rubber, plastic, fabric, reflective object
Movable-static runcones, chocks, barriers, cartsapproved overlay, unapproved temporary, moved between sessions
Sparse LiDAR runthin poles, curbs, low fixturesrange bins, beam dropout, pose jitter
Hard-negative runmarkings, cracks, drains, rubber depositsno FOD and no temporary asset present
Regression runprevious map incidents and near missesexact map/version/config reproduction

Metrics

MetricReport byGate
False deletion rateclass, tile, route, range, weather, LiDAR densityzero unresolved for safety-critical classes
Static preservation rateasset class and localization feature classabove release threshold before publication
FOD evidence retentionobject, size, material, corridorall placed hazardous FOD has retained evidence
Unknown-to-free erroroccluded or unobserved regionzero in planner-consumed layers
Localization degradationroute segment and standno release-zone regression beyond threshold
Reviewer agreementsampled removed objectsdisagreement triggers tile quarantine or relabeling

Acceptance Rules

RuleRationale
A clean visual map is not an acceptance criterion.visual cleanliness can hide false deletion
Removed points are safety evidence and must be archived.incident review needs the evidence that was deleted
FOD-like removals require explicit disposition.temporary and small does not mean safe to erase
Unknown space must remain unknown.map cleaning must not assert current free space
Localization replay is mandatory.preserved points matter because the vehicle must localize
Thresholds are locked before the holdout run.prevents acceptance-set tuning
Quarantined tiles cannot be silently published.unresolved uncertainty must not reach production

Release Artifact

ArtifactContents
False-deletion reportmetric tables, examples, reviewer decisions, unresolved risks
Rejected-layer archiveremoved, restored, and unknown points with provenance
Localization replay reportraw vs cleaned map comparison and failure cases
Semantic diffvector/point-cloud consistency and topology checks
Publication decisionaccept, accept with overlay, quarantine, rerun, or reject

Sources

Public research notes collected from public sources.