3D LiDAR Bounding Box Annotation for Autonomous Driving
How Shaip delivered precise 3D cuboid bounding box annotation across 14 object classes with ~53 boxes per frame on single-frame point cloud data — calibrated against 2D imagery for autonomous vehicle perception at production scale.
Project Overview
As autonomous driving perception moves toward production deployment, the client needed a highly specialized 3D LiDAR annotation pipeline capable of placing precise cuboid bounding boxes at the density required for real AV training data — with multi-attribute labeling and 2D-image calibration.
Shaip built the end-to-end annotation pipeline covering 3D cuboid placement, default size enforcement, multi-attribute labeling, environmental edge-case handling, and reference-image validation — delivering model-ready AV perception datasets across 14 object classes.
Key Status
Object Classes
14
Boxes per Frame
~53
Attribute Layers
5
Occlusion Levels
4
Challenges
- Placing ~53 3D cuboid boxes per frame with precise tight fitting to point cloud edges
- Handling 14 object classes across vehicles, pedestrians, animals, and movable objects
- Working simultaneously with 2D calibration imagery and 3D point cloud data
- Managing ghost point clouds, ground reflections, fog, mist, and water splashes
- Enforcing default size dimensions per class for cross-frame consistency
Solution
3D Cuboid Placement Workflow
Annotators placed precise 3D cuboid bounding boxes on single-frame point cloud data simultaneously with 2D calibration image reference. Each class came with defined default bounding box dimensions to ensure size consistency across frames — cars 3.4–5.1m, trucks 3.4–14m, pedestrians 0.4–1.2m wide.
Multi-Attribute Labeling
Beyond bounding box placement, every annotated object received attribute labels covering vehicle state (driving/parked), two-wheel vehicle state (with/without rider), pedestrian state (standing/sitting/lying), occlusion state (4 levels: none, partial, most, full), and extremities state (protruding objects, open doors, attached equipment).
Environmental Edge-Case Handling
Strict rules governed annotation of heavily occluded or distant objects, ghost point clouds caused by dual LiDAR miscalibration, ground reflections, and environmental noise from water splashes, fog, and mist. Side mirrors, open doors, crane arms, roof racks, and pedestrian luggage were explicitly excluded; emergency sirens and truck bed containers were included.
Tight Box Fit Validation
Boxes were required to be tightly fitted to point cloud edges with minimal visible gaps. Orientation was verified through calibration data and lane context where point clouds were ambiguous. Night and rainy scenes were treated equivalent to daytime conditions for annotation consistency.
Cross-Modal QA
All annotations were validated against reference 2D images to verify correct sizing and orientation. This cross-modal QA layer caught misalignments that would not have been visible from point cloud alone, ensuring production-grade accuracy for downstream AV perception model training.
Project Scope
| Dataset Type | Classes | Boxes/Frame | Attributes | Occlusion Levels | Validation |
|---|---|---|---|---|---|
| 3D LiDAR AV perception | 14 | ~53 average | 5 attribute layers | 4 (0–100%) | 2D image cross-check |
Outcomes
- Established a dense 3D LiDAR annotation pipeline with ~53 boxes per frame at production accuracy
- Standardized 14-class object ontology with default size enforcement per class
- Delivered 5 attribute layers including 4-level occlusion classification
- Implemented cross-modal QA validating 3D boxes against 2D calibration imagery
- Enabled the client’s autonomous vehicle perception, ADAS, and self-driving model training
Overall, Shaip helped transform a complex 3D LiDAR perception requirement into a structured, production-ready annotation pipeline — one capable of supporting autonomous driving, ADAS, robotaxi, and self-driving truck development with calibration-grade accuracy across challenging real-world conditions.
Shaip handled the AV edge cases that break most annotation pipelines — fog, ghost point clouds, dense occlusion, and ~53 boxes per frame. Their 2D-3D cross-validation gave us perception training data we trusted.
– Lead Engineer, Perception Stack
