stefan-jeske/hyve-datasets

Datasets

This directory contains several datasets used for the HYVE project. Please note that each dataset is subject to its own license terms. All datasets are provided in "processed" form, but use the respective dataset as input.

• Armadillo: Sourced from the Stanford 3D Scanning Repository.
• Dragon: Sourced from the Stanford 3D Scanning Repository.
• Objaverse: Sourced from Objaverse (CC0 models).
• Thingi10k: Thingi10k dataset. Sourced from fTetWild, Apache-2.0 licenced from base dataset.

Please verify the LICENSE file in each subdirectory for specific usage rights and restrictions.

The following datasets that were used in the paper are not uploaded here due to license restrictions. Sample model configuration and checkpoints are included in the main repo and code for generating the dataset for the model is available in the main repo under

• Dynamic FAUST: scripts/d_faust_dataset.py
• Shapenet V2 Planes: scripts/shapenetv2_preprocess.py and scripts/shapenetv2_dataset.py
• ScanNet: scripts/scannet_dataset.py

Data Format

Each processed data file is a pytorch-pickled binary .pt file. Each of these files contains data in the following format, where GeometricDataItem actually uses torch_geometric.data.Data objects. The GeometricDataItem simply describes how the data is packed into the format.

class ProcessedDataItem(TypedDict):
    # Data containing both surface and volume samples as input (deprecated and not used anymore)
    volume: GeometricDataItem

    # Data containing only surface samples as input
    surface: GeometricDataItem

    # Unique sample identifier
    id: int

class GeometricDataItem(TypedDict):
    # Positions (x, y, z) of surface (or volume) point samples that should be encoded
    pos: Tensor

    # Point normals for input samples (only used for when specified)
    x: Tensor

    # Data for pre-sampled sdf sample points
    # y[0:3] = sample positions (x, y, z)
    # y[3] = sdf value at sample position (although this may be erroneous due to sampling process)
    # y[4:7] = normal at sample position
    y: Tensor

    # Pair-wise point connectivity, based on k-NN here and typically generated on-the-fly
    batch: Tensor