Dataset¶

Dataset loaders for common 3D benchmarks.

Defined in warpconvnet/dataset/.

ModelNet40Dataset¶

from warpconvnet.dataset.modelnet import ModelNet40Dataset

ModelNet40Dataset(
    root_dir: str = "./data/modelnet40",
    split: str = "train",  # "train" or "test"
)

PyTorch Dataset for ModelNet40 point cloud classification (40 object categories, 2048 points per shape).

Downloads and extracts the HDF5 dataset automatically on first use.

Each sample is a dict:

Key	Shape	Dtype	Description
"coords"	(2048, 3)	float32	Point coordinates
"labels"	scalar	int64	Class label (0-39)

ScanNetDataset¶

from warpconvnet.dataset.scannet import ScanNetDataset

ScanNetDataset(
    root: str = "./data/scannet",
    split: str = "train",          # "train" or "val"
    voxel_size: float | None = None,  # optional voxel downsampling
    out_type: str = "voxel",       # "point" or "voxel"
    min_coord: tuple | None = None,  # optional coordinate offset
)

PyTorch Dataset for ScanNet semantic segmentation using the pre-processed data from the OpenScene project.

Downloads the dataset automatically on first use (~1.3 GB).

Each sample is a dict:

Key	Shape	Dtype	Description
"coords"	(N, 3)	float32	Point coordinates
"colors"	(N, 3)	float32	RGB colors
"labels"	(N,)	int64	Semantic labels (20 classes, 255 = ignore)

If voxel_size is set, points are voxel-downsampled before returning.