The full CoSTAR Block Stacking Dataset is approximately 0.5 TB in size with 16k files. Individual stacking attempt files are typically between 30MB and 50MB, outliers can range from 15kb if there are no images to about 80MB. Text files (txt, csv, yaml, md) with documentation, subsets and configuration information are also included.
The dataset files can be viewed and downloaded individually at the Johns Hopkins CoSTAR Dataset on the Internet Archive. We thank the Internet Archive for graciously hosting our dataset and for their mission of universal knowledge access and preservation! If you have a dataset of your own please consider uploading it to them.
This will most likely be the easiest way for you to get started by downloading the successful stacking examples in the dataset without distractors.
git clone https://github.com/ahundt/costar_dataset.git
cd costar_dataset/costar_dataset
mkdir -p ~/.keras/datasets/costar_block_stacking_dataset_v0.4
python download_success.py ~/.keras/datasets/costar_block_stacking_dataset_v0.4
pip install internetarchive
mkdir -p ~/.keras/datasets/costar_block_stacking_dataset_v0.4
cd ~/.keras/datasets/costar_block_stacking_dataset_v0.4
ia download johns_hopkins_costar_dataset
View other download options:
ia download --help
Download small files including documentation, human created labels, train/val/test subsets, and camera calibration:
mkdir -p ~/.keras/datasets/costar_block_stacking_dataset_v0.4
cd ~/.keras/datasets/costar_block_stacking_dataset_v0.4
ia download johns_hopkins_costar_dataset --glob="*.md"
ia download johns_hopkins_costar_dataset --glob="*.csv"
ia download johns_hopkins_costar_dataset --glob="*.txt"
ia download johns_hopkins_costar_dataset --glob="*.yml"
Files containing the substring 'success':
mkdir -p ~/.keras/datasets/costar_block_stacking_dataset_v0.4
cd ~/.keras/datasets/costar_block_stacking_dataset_v0.4
ia download johns_hopkins_costar_dataset --glob="*success*"
The dataset files can be viewed and downloaded individually at the Johns Hopkins CoSTAR Dataset on the Internet Archive.
Code for this dataset and neural network models is from our paper The CoSTAR Block Stacking Dataset: Learning with Workspace Constraints, with examples and instructions further down on this page:
CoSTAR Dataset Loading Repository — PyTorch and TensorFlow dataset loader
HyperTree Architecture Search Code — aka CoSTAR Hyper, requires TensorFlow
rENAS: regression Efficient Neural Architecture Search Code — requires TensorFlow
CoSTAR Objects - 3D Models of blocks, bin, and other objects
Code for CoSTAR itself is at:
johns_hopkins_costar_dataset_archive.torrent
This will most likely be the fastest way for you to download the dataset. BitTorrent is an efficient distributed data transfer protocol which you can use by installing a bittorrent client application. You then use the small bittorrent file provided by the Internet Archive which tells the client how to download the dataset.
Ubuntu 16.04 download instructions:
# Ubuntu
sudo apt-get update
sudo apt-get install aria2
Then, use the following commands to download the torrent file.
mkdir -p ~/.keras/datasets
cd ~/.keras/datasets
# single line command to perform the full dataset download
aria2c https://archive.org/download/johns_hopkins_costar_dataset/johns_hopkins_costar_dataset_archive.torrent -d ~/.keras/datasets/
After the files are downloaded use the following command to rename the folder.
cd ~/.keras/datasets/
mv johns_hopkins_costar_dataset costar_block_stacking_dataset_v0.4