📝 ⇝ 🧏 Transcription

Repository for sign language transcription related models.

Ideally pose based models should use a shared large-pose-language-model, able to encode arbitrary pose sequence lengths, and pre-trained on non-autoregressive reconstruction.

• _shared - includes shared utilities for all models
• video_to_pose - performs pose estimation on a video
• pose_to_segments - segments pose sequences
• text_to_pose - animates poses using text
• pose_to_text - generates text from poses

Installation

pip install git+git://github.com/sign-language-processing/transcription.git

Development Setup

# Update conda
conda update -n base -c defaults conda

# Create environment
conda create -y --name sign python=3.10
conda activate sign

# Install all dependencies, may cause a segmentation fault
pip install .[dev]

export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python

Example Usage: Video-to-Text

Let's start with having a video file of a sign language sentence, word, or conversation.

curl https://media.spreadthesign.com/video/mp4/13/93875.mp4 --output sign.mp4

Next, we'll use video_to_pose to extract the human pose from the video.

pip install mediapipe # depends on mediapipe
video_to_pose -i sign.mp4 --format mediapipe -o sign.pose

Now let's create an ELAN file with sign and sentence segments: (To demo this on a longer file, you can download a large pose file from here)

pip insatll pympi # depends on pympi to create elan files
pose_to_segments -i sign.pose -o sign.eaf --video sign.mp4

Next Steps

After looking at the ELAN file, adjusting where needed, we'll transcribe every sign segment into HamNoSys or SignWriting:

pose_to_text --notation=signwriting --pose=sign.pose --eaf=sign.eaf

After looking at the ELAN file again, fixing any mistakes, we finally translate each sentence segment into spoken language text:

text_to_text --sign_language=us --spoken_language=en --eaf=sign.eaf

Example Usage: Text-to-Video

Let's start with having a spoken language word, or sentence - "Hello World".

Next Steps

First, we'll translate it into sign language text, in SignWriting format:

text_to_text --spoken_language=en --sign_language=us \
  --notation=signwriting --text="Hello World" > sign.txt

Next, we'll animate the sign language text into a pose sequence:

text_to_pose --notation=signwriting --text=$(cat sign.txt) --pose=sign.pose

Finally, we'll animate the pose sequence into a video:

# Using Pix2Pix
pose_to_video --model=pix2pix --pose=sign.pose --video=sign.mp4 --upscale=true
# OR Using StyleGAN3
pose_to_video --model=stylegan3 --pose=sign.pose --video=sign.mp4 --upscale=true
# OR Using Mixamo
pose_to_video --model=mixamo --pose=sign.pose --video=sign.mp4