Datasets:

anticdimi
/

CloSe-D

license: other
license_name: close-mixed
license_link: LICENSE
pretty_name: CloSe-D
language:
  - en
tags:
  - 3d
  - point-cloud
  - mesh
  - clothing
  - segmentation
  - human
  - smpl
  - 3dv-2024
task_categories:
  - other
task_ids: []
size_categories:
  - 1K<n<10K
configs:
  - config_name: CloSe-Di
    data_files:
      - split: all
        path: CloSe-Di/*.npz
  - config_name: CloSe-Dc
    data_files:
      - split: all
        path: CloSe-Dc/*.npz
  - config_name: CloSe-Dpp
    data_files:
      - split: all
        path: CloSe-D++/*.npz
  - config_name: CloSe-Dpp-updated
    data_files:
      - split: all
        path: CloSe-D++_updated/*.npz

CloSe-D: A 3D Clothing Segmentation Dataset

Project Page · Paper (arXiv:2401.12051) · Code (GitHub)

CloSe-D is a large-scale 3D clothing segmentation dataset accompanying the 3DV 2024 paper "CloSe: A 3D Clothing Segmentation Dataset and Model". It provides per-vertex clothing labels, SMPL body registrations, and canonical-pose annotations for thousands of clothed human scans across 18 garment categories.

Dataset Subsets

The dataset is released in three subsets with different licensing terms.

Subset	# Scans	Scans released?	Labels	SMPL reg.	Notes
CloSe-Di	1,444	Yes	Yes	Yes	Fully releasable
CloSe-Dc	1,732	No (commercial)	Yes	Yes	Purchase scans from Renderpeople, Twindom, AXYZ
CloSe-D++	~1,000	No (original sources)	Yes	—	Labels over THuman2.0, HuMMan, 3DHumans
CloSe-D++_updated	—	No (original sources)	Yes	—	Refined/expanded THuman2.0 labels (`THuman2.0_labels.npz`), released post-paper

For CloSe-Dc, please obtain the raw scans from:

File Format

CloSe-Di / CloSe-Dc — one .npz per scan, keyed by fields such as points, normals, colors, faces, labels, garments, pose, betas, trans, canon_pose, scale. CloSe-Dc omits scan geometry (the commercial scans are not redistributed) and keeps labels + SMPL registration only. See docs/dataset.md for the full field spec.
CloSe-D++ / CloSe-D++_updated — a small number of consolidated .npz files, each acting as a dict of scan_id → label_array. Access via:
```
labels = np.load('CloSe-D++/THumans2.npz')['0000']
```

Garment Label Mapping

{
  0: 'Hat',        1: 'Body',       2: 'Shirt',       3: 'TShirt',
  4: 'Vest',       5: 'Coat',       6: 'Dress',       7: 'Skirt',
  8: 'Pants',      9: 'ShortPants', 10: 'Shoes',      11: 'Hoodies',
  12: 'Hair',     13: 'Swimwear',  14: 'Underwear',  15: 'Scarf',
  16: 'Jumpsuits', 17: 'Jacket'
}

Usage

from huggingface_hub import snapshot_download
import numpy as np

# Download the Di subset
path = snapshot_download(
    repo_id='<user>/CloSe-D',
    repo_type='dataset',
    allow_patterns='CloSe-Di/*',
)

data = np.load(f'{path}/CloSe-Di/<scan_id>.npz')
points, labels = data['points'], data['labels']

For a complete training/inference pipeline (CloSe-Net) and the interactive annotation tool (CloSe-T), see the GitHub repository.

Splits

A reference train/test split for CloSe-Di is provided at splits/split_closedi.npz.

Licensing

The CloSe-D dataset combines material with different licensing terms:

CloSe-Di (scans, labels, SMPL): released by the authors for non-commercial research under MIT-style terms consistent with the accompanying code repository.
CloSe-Dc (labels, SMPL only): the authors release only annotations. The underlying scans are the property of Renderpeople, Twindom, and AXYZ and must be purchased separately under their respective commercial licenses.
CloSe-D++ (labels only): annotations are released by the authors. The underlying scans are subject to the licenses of THuman2.0, HuMMan, and 3DHumans and must be obtained from the original sources.

By downloading this dataset you agree to these terms and to cite the paper below.

Citation

@inproceedings{antic2024close,
    title     = {{CloSe}: A {3D} Clothing Segmentation Dataset and Model},
    author    = {Anti{\'c}, Dimitrije and Tiwari, Garvita and Ozcomlekci, Batuhan and Marin, Riccardo and Pons-Moll, Gerard},
    booktitle = {International Conference on 3D Vision (3DV)},
    month     = {March},
    year      = {2024},
}

If you use CloSe-D++, please additionally cite the source datasets (THuman2.0, HuMMan, 3DHumans) — see the project docs for their BibTeX.

Contact

Dimitrije Antić — d.antic@uva.nl
Garvita Tiwari — gtiwari@mpi-inf.mpg.de