Datasets¶

torch-pointcloud provides several datasets for benchmarking and training. Each dataset returns a single dict (the format consumed by transforms) and integrates with torch.utils.data.DataLoader via the collate helper in torch_pointcloud.utils.data.

Each dataset contains a download parameter (when possible) to automatically download the dataset. Datasets are organized in a raw (containing the raw data) and a processed (containing preprocessed data by torch-pointcloud) directory as follows:

data
├── ModelNet40
│   ├── raw
│   │   ├── airplane
│   │   ├── bathtub
│   │   ├── ...
│   │   └── xbox
│   └── processed
│       ├── train.pt
│       └── test.pt
└── ...

To use them:

from torch.utils.data import DataLoader
from torch_pointcloud.datasets import ModelNet40
from torch_pointcloud.utils.data import collate

dataset = ModelNet40(root="data", train=True, download=True)
dataloader = DataLoader(dataset, batch_size=32, collate_fn=collate)

PointCloudDataLoader is the same DataLoader with the collate helper already applied.

from torch_pointcloud.utils.data import PointCloudDataLoader
from torch_pointcloud.datasets import ModelNet40

dataset = ModelNet40(root="data", train=True, download=True)
dataloader = PointCloudDataLoader(dataset, batch_size=32)

Tasks¶

Classification

ModelNet and ScanObjectNN: meshes, presampled clouds, difficulty variants.
Segmentation

Indoor rooms and outdoor LiDAR, their splits, and the label conventions.
Part segmentation

ShapeNetPart: 16 categories, 50 global part ids.
Detection

SUN RGB-D, KITTI, nuScenes, and batching ragged boxes.

Object classification¶

Dataset	Paper	Samples	Classes
ModelNet10 / ModelNet40	3D ShapeNets: A Deep Representation for Volumetric Shapes	~12k	10 / 40
ModelNet40Hdf5	PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation	~12k	40
ShapeNetPart	A Scalable Active Framework for Region Annotation in 3D Shape Collections	~16k	16 categories / 50 parts
ScanObjectNN	Revisiting Point Cloud Classification: A New Benchmark Dataset and Classification Model on Real-World Data	2.9k	15

Indoor scene segmentation¶

Dataset	Paper	Scenes	Classes
S3DIS	3D Semantic Parsing of Large-Scale Indoor Spaces	271 rooms, 6 areas	13
ScanNet v2	ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes	1.5k scenes	20 (NYU40) / 200

Outdoor / driving segmentation¶

Dataset	Paper	Frames	Classes
SemanticKITTI	SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences	43k frames	19
Semantic3D	Semantic3D.net: A new Large-scale Point Cloud Classification Benchmark	30 scenes	8
Toronto3D	Toronto-3D: A Large-scale Mobile LiDAR Dataset for Semantic Segmentation of Urban Roadways	4 areas	8
ParisLille3D	Paris-Lille-3D: a large and high-quality ground truth urban point cloud dataset for automatic segmentation and classification	3 scenes	9

Base class¶

Dataset	Task	Notes
PointCloudDataset	(any)	Abstract base class all loaders build on: `raw/` + `processed/` disk layout, `download` / `process` hooks. Subclass it for custom data.

About dict keys¶

All datasets emit dicts using the standard key conventions from DataKeys in torch_pointcloud.utils.data:

Key	Shape	Description
`pos`	\((N, 3)\)	3D coordinates
`color`	\((N, 3)\)	RGB (uint8 or float, depending on dataset)
`normal`	\((N, 3)\)	Surface normals (when available)
`segment`	\((N,)\)	Semantic labels (segmentation datasets)
`instance`	\((N,)\)	Instance IDs (ScanNet)
`label`	scalar	Object class (classification datasets)
`face`	\((F, 3)\)	Triangle indices (ModelNet / mesh datasets)

After collate, per-point tensors are concatenated along axis 0 and a batch key of shape \((N,)\) gives each point's source scene.

Color conventions vary per dataset

color is uint8 in \([0, 255]\) for the raw-value loaders (S3DIS, ScanNet, Toronto3D, Semantic3D, SunRGBD) and float32 in \([0, 1]\) for S3DISHdf5, which ships pre-normalized values.

Ignore-index conventions vary per dataset

Unlabeled points use label 0 (<unk> / outdoor conventions, e.g. ScanNet), -1 (indoor no-instance and class-subset remaps, e.g. S3DIS), or 255 (the SemanticKITTI remap example).