README.md

metadata

tags:
  - sign-language
  - how2sign
  - features

How2Sign — Extracted Features

Pre-computed features from the How2Sign dataset.

Layout

Each modality is split by train / test / val and packed into ~3 GB tar shards.

Modality	Content	Approx size
depth_rendered	rendered depth-map JPGs per clip	~39 GB
poses_rendered	rendered pose-skeleton JPGs per clip	~43 GB
poses	raw pose .npy per clip	~18 GB
optical_flow	optical flow .npy per clip	~3 GB
optical_flow_rendered	rendered optical_flow JPGs per clip with stride 2	~11 GB
processed_english_translations	translations csv per split	~6 MB

Inside each tar, paths are relative to the split, e.g.:

<clip_name>/00000.jpg     # rendered modalities
<clip_name>.npy           # npy modalities

Notes on .npy files

Poses .npy is a 0-d object array wrapping a Python object (typically a dict). Load with:

import numpy as np
data = np.load("clip.npy", allow_pickle=True).item()

Downloading

Everything:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="Alexeus17071/How2Sign_with_features",
    repo_type="dataset",
    local_dir="./how2sign_features",
)

Just one modality/split:

snapshot_download(
    repo_id="Alexeus17071/How2Sign_with_features",
    repo_type="dataset",
    local_dir="./how2sign_features",
    allow_patterns=["poses/train/*"],
)

Extract:

mkdir -p extracted/poses/train
for f in how2sign_features/poses/train/*.tar; do
    tar -xf "$f" -C extracted/poses/train/
done