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
| Item | HKCD / PGN3DCD | Airside relevance |
|---|---|---|
| Source paper | PGN3DCD: Prior-Knowledge-Guided Network for Urban 3D Point Cloud Change Detection | Method and dataset for point-level map-change detection |
| DOI | 10.1109/TGRS.2024.3436854 | Peer-reviewed TGRS source |
| Dataset | Hong Kong Change Detection (HKCD) | Realistic urban 3D point-cloud change proxy |
| Scale | About 8.1 km^2 and nearly 128 million annotated points | Comparable scale challenge to large airport surfaces |
| Labels | Binary changed/unchanged point labels | Useful first stage before semantic airport labels |
| Change types | Additions, subtractions/removals, replacements | Maps to new barriers, removed structures, replaced assets |
| Classes discussed | Buildings, terrain, street furniture, trees, vehicles | Airport-specific classes still missing |
| Method | Prior-knowledge-guided 3D change detection network | Shows value of explicit geometric/texture priors |
What Transfers
| HKCD capability | Airside transfer | Required adaptation |
|---|---|---|
| Large bi-temporal point-cloud comparison | Compare previous airport survey tile to current fleet map | Airport ENU frame, survey control, and vehicle-pose uncertainty |
| Point-level binary change masks | Train a changed/unchanged detector | Add semantic labels for FOD, GSE, aircraft, markings, barriers |
| Long-tail class imbalance | Mirrors rare but critical airport changes | Report PR-AUC, mIoU, and safety-weighted recall |
| Added/removed point reasoning | Candidate map patch generation | Occlusion-aware negative evidence before deletion |
| Urban object replacement | Asset replacement or work-zone transition | Operations metadata and human review |
| Public code/dataset page | Reproducible baseline experiments | Confirm dataset access, license, and preprocessing before training |
Benchmark Use
| Experiment | How to run | Airside learning |
|---|---|---|
| Binary changed-point baseline | Train/evaluate PGN3DCD or Siamese point model on HKCD | Establish expected performance on realistic urban changes |
| Add/remove split | Score old-time changed points as removals and new-time changed points as additions | Build separate thresholds for deletion vs insertion |
| Registration stress | Perturb alignment before inference | Quantify false changes from pose/map-frame error |
| Sparse-change stress | Downsample changed points or vary class prevalence | Understand recall under long-tail changes |
| Airport pretraining | Pretrain on HKCD, fine-tune on apron captures | Measure urban-to-airside transfer gap |
| Reviewer workload simulation | Cluster changed points into objects/regions | Estimate number of map QA tickets per airport shift |
Airside Gaps
| Gap | Why it matters |
|---|---|
| No airport apron semantics | Aircraft, tugs, belt loaders, cones, chocks, jet bridges, and FOD have different policy treatment |
| Binary labels only | Airside needs permanent static, movable-static, current dynamic, hazard, artifact, and unknown/review |
| Photogrammetric point clouds | Vehicle LiDAR/radar/camera maps have different density, noise, occlusion, and viewpoints |
| Urban vertical structure bias | Apron changes include low-profile markings, painted surfaces, and small debris |
| No operational status | Airport map updates depend on closures, NOTAM/AIRAC, and sponsor approval |
Recommended Metrics
| Metric | Why use it |
|---|---|
| Changed-point recall | Missing true map changes is the primary safety risk |
| Precision / false positives per hectare | Reviewer burden must remain manageable |
| mIoU changed vs unchanged | Standard point-level benchmark comparison |
| PR-AUC | Handles class imbalance better than accuracy |
| Added vs removed F1 | Insertions and deletions have different operational risks |
| Cluster-level recall | Map QA acts on objects/regions, not isolated points |
| Localization delta | A change detector is only useful if the updated map improves or preserves localization |
Implementation Guidance
- Use HKCD to harden 3D change-detection models before airside data exists, but do not claim airside validity from HKCD alone.
- Preserve added and removed point sets separately. Airport deletion decisions require stronger evidence than insertion alerts.
- Test sensitivity to registration error; false changes from misalignment can dominate real apron changes.
- Convert point masks into object/region proposals before map QA. Human reviewers need clustered features with before/after evidence.
- Fine-tune on airside captures with low-profile markings and small hazards, which are underrepresented in urban building/street-furniture data.
- Add a policy layer after binary change detection so aircraft/GSE and FOD do not flow into the permanent static map.
Sources
- PGN3DCD DOI: https://doi.org/10.1109/TGRS.2024.3436854
- HKCD dataset repository: https://github.com/zhanwenxiao/HKCD
- PGN3DCD code repository: https://github.com/zhanwenxiao/PGN3DCD
- DOI metadata summary: https://colab.ws/articles/10.1109%2Ftgrs.2024.3436854
- Local context: Moved-Object and Map-Change Datasets
- Local context: Airside Dynamic Map Cleaning Benchmark