Week 2: Data Pipeline
Day 8: Capturing Signer Video
Learn what makes source video useful for sign-language AI.
Goal
Learn what makes source video useful.
Learn
- Good capture starts before any model runs. The signer needs visible hands, face, upper body, and enough signing space.
- Use stable framing, clear lighting, consistent distance, high enough resolution, and a background that does not hide hands or face.
- Consent is part of capture quality. Sign-language video can reveal identity, disability status, language community, and personal expression.
Example
- Capture checklist: face visible, both hands visible, wrists not cut off, elbows usually visible, no strong backlight, camera stable, signer centered, no confusing background motion, consent recorded, file name assigned.
- A clip with perfect lighting but no consent is not usable for a responsible dataset.
Practice
- Create a capture checklist with three sections: camera, signer visibility, and consent.
- Use it to grade one existing video as keep, review, or reject.
Checkpoint
Before moving on
You can identify video conditions that help or hurt SignLLM training.
Deaf-first note
Deaf-first note
Do not make signers adapt to a careless pipeline. Build the capture setup around clear signing and respectful consent.