Datasets:

Yong-Hoon
/

dust3r-dataset

license: cc-by-nc-sa-4.0
task_categories:
  - depth-estimation
  - image-to-3d
tags:
  - dust3r
  - 3d-reconstruction
  - stereo-vision
  - pointcloud
  - depth
pretty_name: DUSt3R Preprocessed Training Data
size_categories:
  - 1T<n

DUSt3R Preprocessed Training Data

Preprocessed training datasets for DUSt3R: Geometric 3D Vision Made Easy (CVPR 2024).

These datasets have been preprocessed using the scripts provided in the official DUSt3R repository (datasets_preprocess/) and are ready for training.

Datasets

Dataset	Original Size	Parts	Description	Original License
`arkitscenes_processed`	153 GB	8 (aa~ah)	ARKitScenes - Apple's indoor 3D scene dataset	CC BY-NC-SA 4.0
`blendedmvs_processed`	79 GB	4 (aa~ad)	BlendedMVS - Large-scale multi-view stereo dataset	CC BY 4.0
`co3d_processed`	119 GB	6 (aa~af)	CO3Dv2 - Common Objects in 3D	CC BY-NC 4.0
`habitat_processed`	611 GB	29 (aa~bc)	Habitat-Sim - Photorealistic 3D simulator	See Habitat
`megadepth_processed`	42 GB	3 (aa~ac)	MegaDepth - Single-view depth prediction from internet photos	See MegaDepth
`scannetpp_processed`	36 GB	2 (aa~ab)	ScanNet++ - High-fidelity indoor scene dataset	Non-commercial
`staticthings3d_processed`	43 GB	3 (aa~ac)	StaticThings3D - Synthetic stereo dataset	See StaticThings3D
`waymo_processed`	24 GB	2 (aa~ab)	Waymo Open Dataset - Autonomous driving dataset	Non-commercial
`wildrgbd_processed`	38 GB	2 (aa~ab)	WildRGB-D - Wild RGB-D dataset	See WildRGB-D

Total: ~1.1 TB (59 parts)

File Structure

Each dataset is split into 20 GB parts with the following naming convention:

<dataset_name>.tar.part_aa
<dataset_name>.tar.part_ab
<dataset_name>.tar.part_ac
...
<dataset_name>.sha256          # SHA256 checksums for integrity verification

Download

Download All (bash)

# Using git lfs
git lfs install
git clone https://huggingface.co/datasets/Yong-Hoon/dust3r-dataset

# Or using huggingface-cli
huggingface-cli download Yong-Hoon/dust3r-dataset --repo-type dataset --local-dir ./data

Download a Specific Dataset (bash)

huggingface-cli download Yong-Hoon/dust3r-dataset --repo-type dataset --include "arkitscenes_processed.*" --local-dir ./data

Download All (Python)

from huggingface_hub import snapshot_download

# Download entire dataset
snapshot_download(
    repo_id="Yong-Hoon/dust3r-dataset",
    repo_type="dataset",
    local_dir="./data",
)

Download a Specific Dataset (Python)

from huggingface_hub import snapshot_download

# Download only a specific dataset (e.g., arkitscenes_processed)
snapshot_download(
    repo_id="Yong-Hoon/dust3r-dataset",
    repo_type="dataset",
    local_dir="./data",
    allow_patterns=["arkitscenes_processed.*"],
)

Download Multiple Datasets (Python)

from huggingface_hub import snapshot_download

# Choose datasets to download
datasets_to_download = [
    "arkitscenes_processed",
    "co3d_processed",
    "megadepth_processed",
]

patterns = [f"{name}.*" for name in datasets_to_download]
snapshot_download(
    repo_id="Yong-Hoon/dust3r-dataset",
    repo_type="dataset",
    local_dir="./data",
    allow_patterns=patterns,
)

Checksum Verification

Each dataset includes a .sha256 file for integrity verification.

Verify (bash)

cd data

# Verify a specific dataset
sha256sum -c arkitscenes_processed.sha256

# Verify all datasets
for f in *.sha256; do
    echo "Verifying $f ..."
    sha256sum -c "$f"
done

Verify (Python)

import hashlib
from pathlib import Path


def verify_dataset(data_dir: str, dataset_name: str) -> bool:
    """Verify SHA256 checksums for a dataset."""
    data_path = Path(data_dir)
    sha256_file = data_path / f"{dataset_name}.sha256"

    if not sha256_file.exists():
        print(f"Checksum file not found: {sha256_file}")
        return False

    all_ok = True
    for line in sha256_file.read_text().strip().splitlines():
        expected_hash, filename = line.split()
        filepath = data_path / filename.strip()

        if not filepath.exists():
            print(f"MISSING: {filename}")
            all_ok = False
            continue

        sha256 = hashlib.sha256()
        with open(filepath, "rb") as f:
            for chunk in iter(lambda: f.read(8192 * 1024), b""):
                sha256.update(chunk)

        if sha256.hexdigest() == expected_hash:
            print(f"OK: {filename}")
        else:
            print(f"FAILED: {filename}")
            all_ok = False

    return all_ok


# Verify a specific dataset
verify_dataset("./data", "arkitscenes_processed")

# Verify all datasets
datasets = [
    "arkitscenes_processed", "blendedmvs_processed", "co3d_processed",
    "habitat_processed", "megadepth_processed", "scannetpp_processed",
    "staticthings3d_processed", "waymo_processed", "wildrgbd_processed",
]
for name in datasets:
    print(f"\n=== {name} ===")
    verify_dataset("./data", name)

Decompression

Decompress a Single Dataset

# Merge split parts and extract
cat arkitscenes_processed.tar.part_* | tar xf -

Decompress All Datasets

cd data

for name in arkitscenes blendedmvs co3d habitat megadepth scannetpp staticthings3d waymo wildrgbd; do
    echo "Extracting ${name}_processed ..."
    cat ${name}_processed.tar.part_* | tar xf -
done

Decompress a Specific Dataset to a Custom Directory

cat arkitscenes_processed.tar.part_* | tar xf - -C /path/to/output/

After extraction, the directory structure will be:

data/
  arkitscenes_processed/
  blendedmvs_processed/
  co3d_processed/
  habitat_processed/
  megadepth_processed/
  scannetpp_processed/
  staticthings3d_processed/
  waymo_processed/
  wildrgbd_processed/

License

DUSt3R code: CC BY-NC-SA 4.0 (Naver Corporation)
Each dataset has its own license as listed in the table above. Please make sure to agree to the license of each dataset before use.
This data is provided for non-commercial research purposes only.

Citation

@inproceedings{dust3r_cvpr24,
      title={DUSt3R: Geometric 3D Vision Made Easy},
      author={Shuzhe Wang and Vincent Leroy and Yohann Cabon and Boris Chidlovskii and Jerome Revaud},
      booktitle = {CVPR},
      year = {2024}
}

Acknowledgements

This dataset collection is based on the preprocessing scripts and pair lists provided by the DUSt3R. All original datasets are the property of their respective authors and institutions.