Skip to content

Moved-Object and Map-Change Datasets

Last updated: 2026-05-09

Why It Matters

Map-change detection decides whether the world has changed enough to invalidate a map, trigger a survey, quarantine a tile, or update a route. Moved-object detection is a related but lower-level problem: identify added, removed, or displaced geometry between observations. Both matter for airside autonomy because temporary barriers, parked GSE, aircraft position, construction, lane/stand markings, and FOD can all create discrepancies between the production map and the current apron.

The datasets below cover outdoor HD-map changes, indoor object rearrangement, street-level point-cloud change, simulated stereo V-SLAM change, and urban point-cloud change. They are best used as complementary proxies.

Dataset/Benchmark Table

Dataset / benchmarkSource URLDomain and sensorsLabels / taskBest useMain transfer risk
Argoverse 2 Map Change / Trust but Verify (TbV)https://www.argoverse.org/av2.htmlOutdoor AV logs from six U.S. cities with ring cameras, LiDAR, ego pose, and HD mapsTemporal labels indicate whether a map change is within 30 m of the AV; 1,000 scenarios with 200 real-world HD-map changesHD-map freshness, lane/crosswalk/marking discrepancy detection, map-conditioned perceptionRoad lane changes do not cover apron stand markings, temporary closures, aircraft/GSE occupancy, or FOD
3RScanhttps://github.com/WaldJohannaU/3RScanNaturally changing indoor RGB-D environments, textured meshes, camera poses, semantic instances1,482 reconstructions of 478 environments with object-level alignment and changed-object transformsMoved-object relocalization, object-level change, indoor long-term SLAMIndoor furniture/object changes are useful for reasoning but not for outdoor geodetic maps
3DCDNet / SLPCCDhttps://github.com/wangle53/3DCDNetStreet-level point-cloud change detection derived from SHREC 2021Point-based changed/unchanged learning benchmark with downloadable SLPCCD dataStreet-scene point-cloud change segmentation and neural baseline comparisonStreet point clouds differ from apron geometry, sensor motion, and map-update policy
PPCA-VINShttps://lnexenl.github.io/PPCA-VINS/Unreal Engine environments, stereo cameras, IMU, ground-truth pose, prior point cloudsAdded and removed building point clouds, V-SLAM based point-cloud change detection metrics and baselineLow-cost camera/IMU change detection and controlled ablation for lighting/mirrorsSynthetic buildings and noiseless sensors require real-world validation before map operations use
Urb3DCDhttps://github.com/JorgesNofulla/Point-Cloud-Urban-Change-detectionSimulated urban 3D point clouds for change detection and categorizationBi-temporal urban point-cloud change classes and benchmark scriptsUrban-scale multiclass change detection, synthetic-to-real experiments, point-level mIoUAerial/urban simulation does not capture apron operations or sensor placement on AGVS/GSE

Metrics

MetricApplies toWhat to report
Frame/event precision and recallArgoverse 2 Map Change, airside map-change alertsAlert quality for "map stale near vehicle" at frame and scenario level
F1 / AUROC / PR-AUCBinary map-change or changed-point detectionClass-imbalance-resistant summary with operating threshold
Point-level IoU / mIoUSLPCCD, Urb3DCD, PPCA-VINS, custom point-cloud changePer-class changed/unchanged or added/removed/moved IoU
Added/removed/moved breakdown3RScan, PPCA-VINS, airside map lifecycleSeparate new object, missing object, displaced object, and geometry deformation
Object pose error3RScan-style moved objectsTranslation, rotation, and correspondence success for moved objects
Distance-to-changeHD map and airside validationDistance from detected alert region to ground-truth changed geometry or map element
Localization impactMap operationsScan-to-map residual, false relocalization, covariance, and route-level acceptance before and after a change
Time-to-detectFleet operationsNumber of passes, frames, or hours before a persistent map change is flagged

Airside/Indoor/Outdoor Transfer

Source domainUseful transferAirside gap
AV2 / TbV outdoor roadsHD-map staleness framing, sensor-map cross-checking, temporal alert labelsAirport stand markings, closed areas, temporary cones/barriers, and aircraft/GSE objects are not represented
3RScan indoorObject moved/removed/added reasoning, instance consistency across rescansNo GNSS/INS, no outdoor weather, no vehicle-scale geodetic map
Street-level point cloudsPoint-level changed geometry segmentationRoad furniture and building facades differ from apron surfaces and ground equipment
Synthetic V-SLAMControlled ablations for illumination, mirrors, and sensor costSimulator realism, no airport physics, and no operational traffic
Urban simulated changeMulticlass point-change pipelines and pretrainingMust be retuned for apron object taxonomy and map-layer semantics

Validation Guidance

  1. Separate map-change detection from map update. A high-confidence alert should quarantine or request review before modifying a production map.
  2. Use AV2/TbV to validate HD-map discrepancy models, but build an airside label set for stand markings, stop bars, service roads, aircraft safety envelopes, temporary work zones, and blocked routes.
  3. Use 3RScan-style object alignment ideas for moved GSE and equipment, but add large-object partial observations such as aircraft tails, wings, belt loaders, and buses.
  4. Report "changed but safe" and "changed and safety-critical" separately. A new painted line and a newly parked aircraft should not have the same operational priority.
  5. Validate change persistence across multiple passes, shifts, weather states, and viewpoints before promoting a change into the static map.
  6. Track false positives per km and per stand. A fleet map operation that floods reviewers with harmless alerts will fail operationally even if recall is high.

Sources

Public research notes collected from public sources.