Skip to content

HKCD: Urban 3D Point Cloud Change Detection

Last updated: 2026-05-09

Why It Matters

HKCD is a realistic urban-scale 3D point-cloud change-detection dataset from Hong Kong. It is useful for airport mapping research because it stresses the hard part of change detection: sparse real changes inside a huge mostly unchanged point cloud, with additions, removals, and replacements distributed across urban structures.

It is not an airport dataset. Its value is as a proxy for large-area 3D change labeling, long-tailed change distribution, and point-level evaluation before collecting airside-specific captures.

Dataset Snapshot

ItemHKCD / PGN3DCDAirside relevance
Source paperPGN3DCD: Prior-Knowledge-Guided Network for Urban 3D Point Cloud Change DetectionMethod and dataset for point-level map-change detection
DOI10.1109/TGRS.2024.3436854Peer-reviewed TGRS source
DatasetHong Kong Change Detection (HKCD)Realistic urban 3D point-cloud change proxy
ScaleAbout 8.1 km^2 and nearly 128 million annotated pointsComparable scale challenge to large airport surfaces
LabelsBinary changed/unchanged point labelsUseful first stage before semantic airport labels
Change typesAdditions, subtractions/removals, replacementsMaps to new barriers, removed structures, replaced assets
Classes discussedBuildings, terrain, street furniture, trees, vehiclesAirport-specific classes still missing
MethodPrior-knowledge-guided 3D change detection networkShows value of explicit geometric/texture priors

What Transfers

HKCD capabilityAirside transferRequired adaptation
Large bi-temporal point-cloud comparisonCompare previous airport survey tile to current fleet mapAirport ENU frame, survey control, and vehicle-pose uncertainty
Point-level binary change masksTrain a changed/unchanged detectorAdd semantic labels for FOD, GSE, aircraft, markings, barriers
Long-tail class imbalanceMirrors rare but critical airport changesReport PR-AUC, mIoU, and safety-weighted recall
Added/removed point reasoningCandidate map patch generationOcclusion-aware negative evidence before deletion
Urban object replacementAsset replacement or work-zone transitionOperations metadata and human review
Public code/dataset pageReproducible baseline experimentsConfirm dataset access, license, and preprocessing before training

Benchmark Use

ExperimentHow to runAirside learning
Binary changed-point baselineTrain/evaluate PGN3DCD or Siamese point model on HKCDEstablish expected performance on realistic urban changes
Add/remove splitScore old-time changed points as removals and new-time changed points as additionsBuild separate thresholds for deletion vs insertion
Registration stressPerturb alignment before inferenceQuantify false changes from pose/map-frame error
Sparse-change stressDownsample changed points or vary class prevalenceUnderstand recall under long-tail changes
Airport pretrainingPretrain on HKCD, fine-tune on apron capturesMeasure urban-to-airside transfer gap
Reviewer workload simulationCluster changed points into objects/regionsEstimate number of map QA tickets per airport shift

Airside Gaps

GapWhy it matters
No airport apron semanticsAircraft, tugs, belt loaders, cones, chocks, jet bridges, and FOD have different policy treatment
Binary labels onlyAirside needs permanent static, movable-static, current dynamic, hazard, artifact, and unknown/review
Photogrammetric point cloudsVehicle LiDAR/radar/camera maps have different density, noise, occlusion, and viewpoints
Urban vertical structure biasApron changes include low-profile markings, painted surfaces, and small debris
No operational statusAirport map updates depend on closures, NOTAM/AIRAC, and sponsor approval
MetricWhy use it
Changed-point recallMissing true map changes is the primary safety risk
Precision / false positives per hectareReviewer burden must remain manageable
mIoU changed vs unchangedStandard point-level benchmark comparison
PR-AUCHandles class imbalance better than accuracy
Added vs removed F1Insertions and deletions have different operational risks
Cluster-level recallMap QA acts on objects/regions, not isolated points
Localization deltaA change detector is only useful if the updated map improves or preserves localization

Implementation Guidance

  1. Use HKCD to harden 3D change-detection models before airside data exists, but do not claim airside validity from HKCD alone.
  2. Preserve added and removed point sets separately. Airport deletion decisions require stronger evidence than insertion alerts.
  3. Test sensitivity to registration error; false changes from misalignment can dominate real apron changes.
  4. Convert point masks into object/region proposals before map QA. Human reviewers need clustered features with before/after evidence.
  5. Fine-tune on airside captures with low-profile markings and small hazards, which are underrepresented in urban building/street-furniture data.
  6. Add a policy layer after binary change detection so aircraft/GSE and FOD do not flow into the permanent static map.

Sources

Public research notes collected from public sources.