Tutorials¶

Guided, runnable notebooks covering the library end to end. Each page is rendered from a Jupyter notebook committed under docs/examples/; open it in Colab or download it from the badges at the top of every tutorial.

The three tiers build on each other. Beginner gets a pretrained model running and explains the conventions everything else assumes. Intermediate puts your own data and your own training loop in the middle. Advanced decodes a driving LiDAR sweep into oriented 3D boxes.

Beginner¶

Start here if you have not run a point cloud model before. The quickstart downloads ModelNet40 on first use; the scene tutorial reads a ScanNet scan, which requires accepting the ScanNet terms of use.

Quickstart: classify a point cloud

Load a pretrained model, build a transform pipeline, and classify an object in a few lines.
Segment a scene

Run per-point semantic segmentation on a full indoor scene with a tiling inferer.
Preprocessing pipelines

Compose dict transforms for sampling, normalization, and augmentation, and inspect each step.

Intermediate¶

See how to use the library with custom data and vanilla torch training loop.

Use your own data

Wrap your own point cloud files in a dataset with disk caching and packed-batch collation.
Train a model

Train a classification model from scratch: dataloaders, optimizer, and the evaluation loop.
Features and similarity search

Read a frozen encoder's features, query one point, and retrieve whole shapes by descriptor.

Advanced¶

See how to use this library for pratical challenges like 3D object detection.

Detect objects in driving LiDAR

Turn one sweep into oriented 3D boxes: voxel encoding, decoding, non-maximum suppression.