Industry Research

Appearance

3D Point Cloud Annotation Tools and Workflows ​

Practical Guide for Creating Airside Driving Datasets ​

Last updated: 2026-05-24

1. The Annotation Challenge for Airside ​

You need 3D bounding box annotations for objects never seen in standard driving datasets:

Object ClassSize (L×W×H m)Challenge
Aircraft (A320)37.6 × 35.8 × 11.8Enormous, partially visible, dynamic wings
Baggage tractor3.5 × 1.5 × 1.8Similar to cars but different shape
Belt loader8.0 × 2.5 × 3.5Articulated, changes shape when deployed
Pushback tug6.0 × 2.5 × 2.0Low profile, hard to distinguish
ULD container3.2 × 2.4 × 1.6Rectangular, stacked, on dollies
Ground crew0.5 × 0.5 × 1.8Small, hi-vis clothing
Baggage dolly train12+ × 1.5 × 1.0Multiple connected units
Fuel truck8.0 × 2.5 × 3.0Hose deployment changes footprint
Catering truck7.0 × 2.5 × 4.0Scissor lift changes height
FOD0.05-0.3 × 0.05-0.3 × 0.05Tiny, ground-level

Estimated annotation speed: 20-40 3D boxes per hour per annotator (experienced). Airside objects are harder than cars — unusual shapes, varying configurations.

2. Open-Source 3D Annotation Tools ​

2.1 CVAT (Computer Vision Annotation Tool) ​

Best for: Teams, multi-format support, integration with auto-labeling

Setup:
  docker-compose up -d  # Self-hosted
  # Or use app.cvat.ai (cloud, free tier available)

3D Support:
  - 3D cuboid annotation on point clouds (Velodyne, custom formats)
  - Supports PCD, PLY, BIN point cloud formats
  - Timeline scrubbing for sequential frames
  - Interpolation between keyframes (reduces annotation by ~5x)

Strengths:
  + Most mature open-source option
  + Team collaboration (multi-user, review workflows)
  + REST API for automation
  + Active development (Intel)

Limitations:
  - 3D UI can be slow with large point clouds (>100K points)
  - No native support for multi-LiDAR merged clouds
  - Limited 3D auto-annotation integration

2.2 Label Studio ​

3D Support Status: Limited
  - 3D point cloud annotation via plugin (community)
  - Not as mature as CVAT for 3D
  - Better for image/text annotation

Verdict: Use CVAT for 3D, Label Studio for 2D camera annotation

2.3 SUSTechPOINTS ​

Best for: Research, fast 3D annotation

GitHub: github.com/naurril/SUSTechPOINTS
Setup: Web-based, runs locally

Features:
  - Purpose-built for LiDAR point cloud annotation
  - Supports BIN/PCD formats
  - Auto-annotation via pre-trained models
  - Ground removal tool
  - One-click box fitting from point selection
  - Frame interpolation

Verdict: Good for small-scale research annotation. Less team features than CVAT.

2.4 3D-BAT (3D Bounding Box Annotation Tool) ​

GitHub: github.com/walzimmer/3d-bat
Focus: Multi-sensor (camera + LiDAR) synchronized annotation
Feature: Project 3D boxes onto camera images for verification
Verdict: Useful when you add cameras and need cross-modal verification

2.5 scalabel ​

GitHub: github.com/scalabel/scalabel (Berkeley)
Focus: Designed for BDD100K driving dataset creation
Features: 3D box tracking, category hierarchy, export to multiple formats
Verdict: Good alternative to CVAT with driving-specific features

3. Commercial Annotation Services ​

Service3D LiDAR SupportPricingTurnaroundQuality
Scale AIYes (primary focus)$0.50-2.00/box24-72 hoursHigh (ML-assisted + human)
LabelboxYesCustomVariesHigh
V7Yes (Darwin)$0.30-1.50/box24-48 hoursMedium-High
Segments.aiYes$0.20-1.00/boxVariesMedium
SuperviselyYes$0.40-1.50/box24-72 hoursHigh
KognicYes (automotive-focused)CustomCustomVery High (AV-specific)

For airside: Consider Kognic or Scale AI — they understand AV annotation requirements. Budget ~$1-2 per 3D box for quality annotations.

Cost estimate for initial airside dataset:

1,000 frames × 15 objects/frame = 15,000 boxes
At $1.50/box = $22,500
Or: self-annotate with auto-labeling pipeline (slower, cheaper)

4. Auto-Labeling Pipeline ​

4.1 The Active Learning Loop ​

Iteration 0: Run nuScenes-pretrained CenterPoint on airside data
  → Noisy labels (wrong classes, missed objects, false positives)
  → Human corrects 200-500 frames (1-2 days of annotation)

Iteration 1: Fine-tune CenterPoint on corrected 200-500 frames
  → Better labels (correct classes, fewer missed objects)
  → Human corrects another 200-500 frames (focus on hard cases)

Iteration 2: Fine-tune on accumulated 400-1000 frames
  → Good labels (most objects detected correctly)
  → Human corrects only the mistakes (~100 frames)

Iteration 3+: Continue until quality plateaus
  → Typically 3-5 iterations, total 500-1500 manually corrected frames

4.2 Auto-Label Quality Tiers ​

python
def tier_detections(detections, thresholds):
    """Sort auto-labels into quality tiers for review prioritization."""
    tiers = {
        'auto_accept': [],    # score > 0.85, use without review
        'spot_check': [],     # 0.6 < score < 0.85, review 10%
        'human_review': [],   # 0.3 < score < 0.6, review all
        'reject': [],         # score < 0.3, discard
    }

    for det in detections:
        if det['score'] > 0.85:
            tiers['auto_accept'].append(det)
        elif det['score'] > 0.6:
            tiers['spot_check'].append(det)
        elif det['score'] > 0.3:
            tiers['human_review'].append(det)
        else:
            tiers['reject'].append(det)

    return tiers

# Expected distribution after Iteration 2:
#   auto_accept: ~60% of detections
#   spot_check: ~20%
#   human_review: ~15%
#   reject: ~5%
# Total human effort: review ~35% of detections (vs 100% from scratch)

4.3 Foundation Model-Assisted Annotation (When Cameras Available) ​

Pipeline:
  1. Camera image → Grounding DINO (text prompt: "baggage tractor")
     → 2D bounding boxes with class labels

  2. SAM/SAM 2 → instance segmentation masks from Grounding DINO boxes

  3. LiDAR point cloud → project into camera frame
     → Select LiDAR points within SAM mask

  4. Selected LiDAR points → fit 3D bounding box
     → L-shape fitting or minimum bounding box

  5. Human reviews 3D box → adjusts size, position, heading

This reduces annotation from:
  "Draw a 3D box from scratch" (60-90 seconds/object)
  to: "Verify and adjust a pre-fitted box" (10-20 seconds/object)
  = 4-6x speedup

4.4 Map-Scale Semantic Candidate Review ​

For aggregated LiDAR maps, do not route open-vocabulary outputs through the same confidence tiers as closed-set 3D boxes. VESPA, ZOPP, SALT, SAM4D, OpenUrban3D, UniLiPs, and LOSC can pre-fill tiles, superpoints, object proposals, or back-projected scan labels, but annotation tasks should show them as read-only pre-annotations with candidate_concept, prompt, model, calibration, consolidation, and tile provenance. Reviewers map candidates to the controlled taxonomy, request a taxonomy change, retain unknown, or reject. Export only qa_passed semantic patches and approved back-projected labels; prompt strings must not become class IDs directly.

For S2+ use, the pre-labeler should resolve to an offboard labeler registry record (../mlops/offboard-labeler-registry-by-scale.md) that names the prompt pack, model/checkpoint, thresholds, retrieval corpus, projection code, allowed-use state, reviewer workflow, QA report, and rollback bundle. Annotation tools can accelerate review, but they should not hide which machine generated the pre-label.

4.5 Annotation Workflow by MLOps Scale ​

Tool choice should follow the maturity of the label workflow. A small research batch can use a local point-cloud tool, but S3-S5 operations need reviewer assignment, schema locks, QA sampling, audit trails, vendor metrics, and promotion states that downstream MLOps can enforce.

MLOps scaleTooling postureRequired workflow controlsMain risk
S0 notebook researchLocal SUSTechPOINTS, CVAT project, or scripted pseudo-label viewerSave data pointer, schema, and examples that explain the resultLabels cannot be reconstructed
S1 repeatable prototypeShared CVAT/Scalabel project with fixed instructionsDataset snapshot, class schema, reviewer notes, export formatAnnotators silently change conventions
S2 production productTeam annotation platform integrated with dataset manifests and model registryPre-label provenance, task status, QA sampling, rejected-label trackingPre-labels leak into training as ground truth
S3 fleet and multi-siteMulti-site task queues, active-learning scores, and local holdout protectionSite/ODD tags, reviewer agreement, duplicate control, budget ledgerCommon sites consume budget while rare local hazards remain unlabeled
S4 regulated safety-criticalEvidence-grade workflow for hazards, FOD, personnel, protected zones, and map-release statesExpert review, immutable audit, approval identity, replay/safety-case linksEvidence labels cannot support an audit or incident review
S5 platform scaleShared annotation service with vendor management, privacy tiers, and policy APIsCross-team schemas, cost attribution, role-based approvals, retention/export policyTeams fork taxonomies and cannot reuse labels safely

For airside LiDAR and image labeling, the annotation interface should always show enough context to prevent false static labels: temporal frames, ego pose, calibration/projection quality, map tile state, reviewer disposition, and whether an object is permanent, movable, dynamic, transient, FOD, artifact, or unknown.

5. Annotation Format Standards ​

5.1 3D Bounding Box Specification ​

A 3D bounding box is defined by 7 parameters:
  (x, y, z, dx, dy, dz, heading)

  x, y, z: Center of the box in vehicle coordinates [meters]
  dx: Length (along heading direction) [meters]
  dy: Width (perpendicular to heading) [meters]
  dz: Height [meters]
  heading: Yaw angle [radians], measured from x-axis (forward)

Vehicle coordinate system (matching RoboSense/nuScenes):
  x: forward
  y: left
  z: up
  Origin: rear axle center

5.2 Format Comparison ​

FormatUsed ByBox RepresentationExtras
KITTIKITTI benchmark(h,w,l,x,y,z,ry) — camera frame!Truncation, occlusion levels
nuScenesnuScenes devkit(x,y,z,w,l,h) + quaternionVelocity, visibility, attributes
WaymoWaymo OD(cx,cy,cz,l,w,h,heading)Speed, acceleration
OpenLABELASAM standardJSON-LD, flexibleMost interoperable
OpenPCDetOpenPCDet framework(x,y,z,dx,dy,dz,heading)Simple, direct

Recommendation: Use OpenPCDet format for training (simplest), export to nuScenes format for sharing/benchmarking.

6. Airside Annotation Guidelines ​

6.1 Class Taxonomy ​

yaml
classes:
  # Primary (always annotate)
  - aircraft          # Any aircraft, regardless of type
  - baggage_tractor   # Tow tractor for baggage carts
  - belt_loader       # Conveyor belt loader
  - pushback_tug      # Aircraft pushback vehicle (towbar or towbarless)
  - ground_crew       # Airport ground personnel (any role)
  - uld_container     # Unit Load Device (on dolly or standalone)
  - trailer           # Baggage/cargo trailer (single unit)

  # Secondary (annotate when present)
  - fuel_truck        # Aircraft refueling vehicle
  - catering_truck    # Catering/provisioning vehicle (scissor lift)
  - maintenance_vehicle # Maintenance/support vehicle
  - fire_truck        # Airport rescue and firefighting vehicle
  - follow_me_car     # Follow-me/marshalling vehicle
  - de_icing_truck    # De-icing/anti-icing vehicle

  # Optional
  - dolly_train       # Connected string of baggage dollies
  - gpu               # Ground Power Unit
  - passenger_stairs  # Mobile passenger stairway
  - apron_bus         # Passenger transfer bus
  - fod               # Foreign Object Debris (ground-level)

6.2 Annotation Rules ​

1. ANNOTATE if: Object is within 80m range AND has ≥ 10 LiDAR points
2. SKIP if: Object is beyond 80m OR has < 10 points (too sparse)
3. OCCLUDED objects: Annotate if ≥ 30% visible, mark occlusion level (0-3)
4. AIRCRAFT: Use tight box around fuselage only (not wings extending beyond sensors)
   - Mark aircraft_type attribute if identifiable (A320, B737, etc.)
5. DOLLY TRAINS: Annotate each dolly separately, not the whole train as one box
6. GROUND CREW: Annotate even if small — safety-critical class
7. FOD: Only annotate if clearly visible (≥ 5 points, ground-level, < 0.5m height)
8. HEADING: Point in the direction of travel (or forward direction for stationary vehicles)

6.3 Quality Metrics ​

Inter-Annotator Agreement (IAA):
  IoU threshold: 0.5 for position, 0.7 for size
  Heading tolerance: ±15°
  Target IAA: > 0.85 for primary classes, > 0.75 for secondary

Quality checks per batch:
  - Random sample 10% of frames for double-annotation
  - Compare IAA metrics
  - Flag annotators below threshold for retraining
  - Flag frames with > 30% disagreement for expert review

7. Recommended Workflow ​

Phase 1: Bootstrap (Week 1-2)
├── Run CenterPoint (nuScenes pretrained) on 500 frames
├── Export detections as initial labels
├── Correct 200 frames in CVAT (primary classes only)
├── Fine-tune CenterPoint on corrected data
└── Deliverable: First airside detector, ~40% mAP

Phase 2: Iterate (Week 2-4)
├── Auto-label remaining frames with improved model
├── Correct another 300 frames (focus on missed objects)
├── Add secondary classes
├── Fine-tune again
└── Deliverable: Improved detector, ~55% mAP

Phase 3: Scale (Week 4-8)
├── Process all available bag data through auto-labeler
├── Spot-check auto-labels (10% review rate)
├── Add camera-assisted annotation when cameras available
├── Build evaluation set (100 frames, exhaustively annotated)
└── Deliverable: Production-quality detector, ~65%+ mAP

Phase 4: Continuous (Ongoing)
├── Active learning: mine hard examples from fleet
├── Human annotates model failures
├── Retrain monthly
└── Target: >75% mAP with < 5% review rate

Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim