Download Datasets

Both LSFB datasets, i.e. cont and isol, can be downloaded through HTTP/HTTPS. The lsfb-dataset package provide a Downloader class taking care of the download according to your needs. For example, the downloader can filter out the files that you don't need and can also resume the downloading where it stops.

Name of the dataset	ID	Poses	Videos (GB)
LSFB ISOL	isol	10GB	25GB
LSFB CONT	cont	31GB	~400GB

As you can see in this table, the datasets can be heavy, especially the videos of the LSFB CONT dataset.

Basic Usage

Download LSFB ISOL Landmarks and Videos

By default, the downloader will fetch the landmarks of the entirety of the specified dataset. The only mandatory parameters are the dataset name and the destination folder where the files are going to be downloaded.

from lsfb_dataset import Downloader

downloader = Downloader(dataset='isol', destination="./destination/folder", include_videos=True)
downloader.download()

Download LSFB CONT Landmarks and Videos

By default, the downloader will fetch the landmarks of the entirety of the specified dataset. The only mandatory parameters are the dataset name and the destination folder where the files are going to be downloaded.

from lsfb_dataset import Downloader

downloader = Downloader(dataset='cont', destination="./destination/folder", include_videos=True)
downloader.download()

A more complex example

Here's a more complex example where we only download the instances: * Of the subsets train, fold_0 and fold_2; * Only the instances of the signers 20 to 39; * Only download the raw poses (without any interpolation of the missing landmarks nor smoothing); * Only includes the landmarks of the pose (body) and the hands; * Without skipping the existing files.

from lsfb_dataset import Downloader

downloader = Downloader(
    dataset='isol',
    destination="./destination/folder",
    splits=['train', 'fold_0', 'fold_2'],
    signers=list(range(20, 40)),
    include_cleaned_poses=False,
    include_raw_poses=True,
    include_videos=False,
    landmarks=['pose', 'left_hand', 'right_hand'],
    skip_existing_files=False,
)
downloader.download()