Datasets

DeePoints comes with the possibily of easily training your models on either segmentation or classification datasets.

# import all builtin datasets
from deepoints.datasets import *

Note

We deliberately decided to define the datasets extending the lightning.LightningDataModule interface. Why? This becomes handy when also adopting the lightning.Trainer class to train any architecture in a more organized and concise way.

Segmentation

Each dataset for semantic segmentation tasks has a classification label assigned to each single point within the cloud. Point features usually include x, y, z, intensity, red, blue and green. We provide the following segmentation-oriented datasets.

• Semantic3D as a dataset of over four billion points and nine classes.

Classification

Differently from segmentation, each point cloud is mapped to a single class label defining the three-dimensional scene while all other point features do not change.

TODO