v2026

2026 Model Release

Improved sign language segmentation model trained on DGS Corpus 3.0.0, using a CNN + RoPE Transformer architecture.

Model

• Architecture: Two-stage UNet CNN spatial encoder + 4-layer pre-norm Transformer with Rotary Position Embedding (RoPE)
• Checkpoint: bundled in the pip package (dist/2026/best.ckpt, bfloat16, 11.5MB)
• Test set (DGS Corpus): Sign IoU = 0.652, Phrase IoU = 0.925, HM = 0.764

Usage

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

wget -O example.pose https://datasets.sigma-sign-language.com/poses/holistic/dgs_corpus/1413451-11105600-11163240_a.pose
pose_to_segments --pose example.pose --elan output.eaf

Server

docker build -t segmentation-serve .
docker run -p 8080:8080 -e PORT=8080 segmentation-serve

# JSON API (signs and sentences in seconds)
curl "http://localhost:8080/?pose=/path/to/input.pose"

# Save ELAN file
curl -X POST http://localhost:8080/ \
  -H "Content-Type: application/json" \
  -d '{"input": "/path/to/input.pose", "output": "/path/to/output.eaf"}'

See dist/2026/README.md for full details on architecture, training, and experiments.

v2023.0.3

Full Changelog: v2023.0.2...v2023.0.3

v2023.0.2: Add health endpoint

feat(server): add health endpoint

v2023.0.1: Release on docker

fix(docker): only install torch cpu

Linguistically Motivated Sign Language Segmentation (2023)

Code for the paper Linguistically Motivated Sign Language Segmentation.