Skip to content

Airside Map Hygiene Ground Truth Protocol

Last updated: 2026-05-09

Map hygiene validation needs ground truth that separates permanent static structure, movable-static assets, current dynamic actors, FOD/hazards, artifacts, and unknown space. Without that separation, a benchmark can reward a cleaner for deleting the very evidence needed for safety review.

Ground Truth Goals

GoalPractical requirement
Measure false deletionlabels cover static assets and FOD-like hazards, not only moving actors
Measure ghost retentiondynamic and transient traces are labeled in accumulated maps
Support release reviewlabels are linked to map tile, route, stand, source frames, and reviewer
Support auditabilityraw evidence and rejected/kept decisions can be traced after publication
Support transfer analysispublic benchmark classes map to airside-specific classes

Label Taxonomy

LabelDefinitionExamplesMap use
permanent_staticapproved, persistent site structureterminal edge, pole, curb, fixed cabinet, stand paintretain
safety_staticstatic feature tied to rules or marginsstop bar, service-road edge, restricted arearetain and validate semantically
movable_staticstationary now but not permanentcone, chock, barrier, parked GSE, cartoverlay or review
dynamic_actormoving during capture or operationally transientaircraft, tug, bus, worker, vehicleremove from permanent map
fod_hazarddebris or small unsafe objectbolt, strap, rubber, plastic, toolhazard/review layer
artifactsensor or registration artifactmultipath, rain/spray, ghost, packet issueremove or mark diagnostic
unknowninsufficient evidenceoccluded, sparse, ambiguous, unobservedquarantine or review

Capture Plan

CapturePurposeMinimum slices
quiet surveyhigh-confidence static referencelow traffic, slow pass, full route/stand coverage
busy operationdynamic and movable-static clutteraircraft present, GSE staged, personnel/vehicles nearby
change pairdistinguish moved assets from permanent structureobject present and absent across sessions
FOD placementsmall-hazard retentioncontrolled articles by size/material/location
sparse/degradedweak-observation evidencerange bins, beam dropout, night/wet if in ODD
hard negativefalse alert and false deletion controlmarkings, cracks, drains, rubber deposits, shadows

Annotation Requirements

FieldRequirement
object_idstable ID across frames, map layers, and review exports
label_classone taxonomy class plus optional subclass
geometry3D cuboid, polygon, point cluster, semantic mask, or map cell set
coordinate_systemmap frame plus sensor frame transforms where labels originate
temporal_extentfirst/last observation, session ID, permanence evidence
source_evidenceraw frame IDs, image crops, LiDAR cluster IDs, reviewer notes
confidencelabel confidence and reason for unknown/review
dispositionretain, remove, overlay, hazard alert, quarantine, or ignore

ASAM OpenLABEL is a good exchange format because it supports multi-sensor labels, coordinate systems, object annotations, scenario tags, and extensible taxonomies. Use an airport-specific ontology for classes that public road datasets do not cover.

QA And Split Rules

RuleRationale
Double-label safety-critical static assets and FOD.false deletion claims need high label quality
Keep acceptance zones separate from tuning zones.prevents threshold overfit
Label unknown explicitly.unobserved space must not become assumed free space
Preserve reviewer disagreement.disagreement identifies taxonomy or evidence gaps
Include site-specific hard negatives.public datasets miss local pavement, lighting, and equipment
Version labels with map and cleaner releases.changing ground truth changes acceptance history

Acceptance Outputs

OutputConsumer
labeled static and dynamic point setsmap-cleaning metric pipeline
FOD/hazard ground truthsafety case and FOD workflow
movable-static inventorymap operations lifecycle
unknown/quarantine polygonspublication gate
annotation manifestaudit, replay, and regression
benchmark split manifestrepeatable release testing

Sources

Public research notes collected from public sources.