The open dataset
Every approved contribution is released openly so researchers and other Deaf-serving organizations can build on it instead of starting from nothing.
- 0
- Clips
- 0
- Signs
- 0
- Contributors
- 0
- Languages
What is in it
Each approved clip ships with:
- The video as recorded, plus a poster frame.
- Raw landmarks — MediaPipe Holistic output, 1662 floats per frame (pose 33x4, face 468x3, both hands 21x3).
- A processed tensor — the same clip time-normalized to 50 frames of 516 body-relative features, ready to train on.
- Quality metrics — frame count, estimated FPS, hand visibility ratio and motion score.
Licence and consent
Released under CC BY 4.0. Every clip was contributed deliberately by someone who agreed to its inclusion — nothing here was scraped. Contributors can withdraw from future releases, so treat each release as a snapshot rather than a permanent archive, and honour removal requests when you redistribute.
Access
Programmatic access for training runs is available through an API key — email support@softbrane.com describing what you are building.