phoenix14t-features / README.md
Shakibyzn's picture
update data fields
b060cd4 verified
metadata
language:
  - de
  - dgs
tags:
  - sign-language
  - phoenix14t
  - i3d
  - s3d
  - vit
license: mit

Phoenix14T Extracted Features

This dataset contains pre-computed visual features for the Phoenix14T German Sign Language (DGS) dataset. These features are extracted using I3D, S3D, and CLIP ViT models.

Dataset Description

Original Dataset: RWTH-PHOENIX-Weather 2014 T
Models Used:

  • I3D: Inflated 3D ConvNet
  • S3D: Separable 3D ConvNet.
  • CLIP (ViT-B/32 or L/14)

Data Structure

The dataset is organized by model type. Each entry corresponds to a video segment from the Phoenix14T train, dev, or test splits.

Feature Dimensions

Model Feature Dimension
I3D 1024
S3D 832
CLIP (ViT) 768

Data Fields

  • name: The original Phoenix14T clip name (e.g., 01January_2011_Friday_tagesschau_default-0).
  • sign: The extracted tensor (NumPy array).
  • gloss: The corresponding sign language glosses.
  • text: The German translation.

Usage

from huggingface_hub import snapshot_download

# Load a specific feature set
snapshot_download(repo_id="Shakibyzn/phoenix14t-features", allow_patterns="data/i3d/*.npz", repo_type="dataset")