You can download the data used in our paper directly. Or you can try to reproduce the data from scratch.
The final data structure of datasets/
is as follows:
datasets
├── data_transform.py
├── nr3d
│ ├── nr3d.csv (optional)
│ └── nr3d_sa.csv
├── README.md
├── scannet -> /PATH/TO/SCANNET
│ ├── instruct-insertion
│ ├── scannetv2-labels.combined.tsv
│ ├── scans
│ ├── scans_test
│ └── tasks
└── sr3d
│ ├── sr3d+.csv (optional)
└── sr3d+_sa.csv
First, go to ScanNet Repo and get their downloading script as instruct-insertion/scripts/download_scannet.py
. We did not include the complete script in our repo because ScanNet is released under their Term of Use, not MIT license.
We modified the script to remove unnecessary data. Make sure the following lines in instruct-insertion/scripts/download_scannet.py
:
FILETYPES = [
".aggregation.json",
# ".sens",
".txt",
# "_vh_clean.ply",
"_vh_clean_2.0.010000.segs.json",
"_vh_clean_2.ply",
# "_vh_clean.segs.json",
# "_vh_clean.aggregation.json",
"_vh_clean_2.labels.ply",
# "_2d-instance.zip",
# "_2d-instance-filt.zip",
# "_2d-label.zip",
# "_2d-label-filt.zip",
]
Since the ScanNet dataset is large, we recommend you to create datasets/scannet
as a soft link to somewhere in the HDD before running our downloading script.
python instruct-insertion/scripts/download_scannet.py -o datasets/scannet
After the ScanNet is downloaded, preprocess the data by running:
bash instruct-insertion/scripts/preprocess_scannet_data.sh
The Sr3D-SA
and Nr3D-SA
datasets can be downloaded from the Google Drive.
Put them under datasets/sr3d
and datasets/nr3d
respectively.
If you want to build the data from source, you should set up ReferIt3D dataset and OpenAI API Token first.
Download ReferIt3D dataset:
pushd datasets
# Download Nr3D
gdown https://drive.google.com/file/d/1qswKclq4BlnHSGMSgzLmUu8iqdUXD8ZC/view?usp=drive_link -O nr3d/nr3d.csv
# Download Sr3D+
gdown https://drive.google.com/file/d/1kcxscHp8yA_iKLT3BdTNghlY3SBfKAQI/view?usp=drive_link -O sr3d/sr3d+.csv
popd
Check datasets/data_transform.py for more details.