What is 3d object detection?
Unlike 2D detection, this task predicts boxes with position, size and orientation in 3D. It supports road-user inventories and asset candidates, depending on training labels.
Approaches and starting model
Point-based, voxel-based and pillar-based detectors are common approaches. PointPillars is the starting example here; its KITTI checkpoint covers road users, not every infrastructure asset.
Inputs and outputs
Prepared LiDAR coordinates and the configured point features enter the network. Boxes contain a position, dimensions and heading, with class and confidence. Match point range, coordinate axes and intensity conventions to the selected configuration.
Where it helps in GIS and digital twins
Useful as a starting point for mobile-mapping asset review. Transform accepted boxes into the project's coordinate frame, attach acquisition time and confidence, and keep a reviewer-approved link to the corresponding asset.
How to start
Follow OpenPCDet's dataset preparation and training instructions with the pointpillar configuration. For another asset type, annotate oriented 3D boxes and change the class configuration before training. Split independent survey routes rather than neighboring frames.
How to evaluate results
Measure class-specific 3D detection performance on held-out local labels. Review false positives, dimensions and heading. Do not turn low-confidence boxes directly into an authoritative asset register.
Limitations
Sparse distant returns, occlusion and different sensors can degrade results. Road-user weights do not identify buildings or utilities. A bounding box is not an exact mesh or asset identity.
Official starting-model documentation
The category artwork illustrates its starting model; it is not an evaluated prediction.
