README.md

metadata

license: other
license_name: research-only
license_link: https://3dmatch.cs.princeton.edu/
task_categories:
  - other
tags:
  - point-cloud
  - 3d-registration
  - 3dmatch
  - geometric-features
  - fcgf
  - correspondence
pretty_name: FCGF Preprocessed 3DMatch
size_categories:
  - 1K<n<10K

FCGF Preprocessed 3DMatch Dataset

Preprocessed 3DMatch training data used by FCGF: Fully Convolutional Geometric Features (ICCV 2019).

Each .npz file stores a fragment point cloud, and the accompanying .txt files list overlapping fragment pairs (with their overlap ratio) used to sample positive correspondences during training. This is the exact data produced by scripts/download_datasets.sh in the FCGF repository.

Contents

threedmatch/
├── <scene>@seq-XX_YYY.npz          # 2189 fragment point clouds
└── <scene>@seq-XX-<overlap>.txt    #  401 overlapping-pair lists

• 2,590 files total (2,189 .npz + 401 .txt), ~8.2 GB.
• Scenes are drawn from the standard 3DMatch compilation: 7-scenes, sun3d, bundlefusion, rgbd-scenes-v2, analysis-by-synthesis, etc.

Usage

Download with the Hugging Face CLI:

hf download chrischoy/FCGF-3DMatch --repo-type dataset --local-dir ./data
# data/threedmatch/*.npz

or from Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="chrischoy/FCGF-3DMatch",
    repo_type="dataset",
    local_dir="./data",
)

Then train FCGF:

python train.py --threed_match_dir ./data/threedmatch/

A single fragment file can be inspected with NumPy:

import numpy as np

data = np.load("data/threedmatch/7-scenes-chess@seq-02_000.npz")
print(data.files)          # e.g. ['pcd', ...]
xyz = data["pcd"]          # (N, 3) point coordinates

License & attribution

This is a redistribution of preprocessed data derived from the 3DMatch benchmark, which itself aggregates several RGB-D datasets (SUN3D, 7-Scenes, BundleFusion, RGB-D Scenes v2, and others). It is provided for non-commercial research purposes only. Please also comply with the licenses of the original constituent datasets and cite 3DMatch. The FCGF source code is released separately under the MIT License.

Citation

If you use this data, please cite FCGF and 3DMatch:

@inproceedings{FCGF2019,
    author    = {Christopher Choy and Jaesik Park and Vladlen Koltun},
    title     = {Fully Convolutional Geometric Features},
    booktitle = {ICCV},
    year      = {2019},
}

@inproceedings{zeng20163dmatch,
    author    = {Andy Zeng and Shuran Song and Matthias Nie{\ss}ner and
                 Matthew Fisher and Jianxiong Xiao and Thomas Funkhouser},
    title     = {3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions},
    booktitle = {CVPR},
    year      = {2017},
}