Week 3: Training and Review
Day 16: Annotation and Labeling
Goal
Learn why labels are expensive and important.
Learn
- Annotation may include gloss, translation, start and end frames, signer information, hand dominance, facial markers, role shift notes, confidence flags, and QA decisions.
- Good labels often require people who know the language. A non-fluent labeler may miss grammar, naturalness, or meaning.
- Annotation should be versioned. If a gloss convention changes, the dataset should show which samples were updated.
Example
- Annotation form fields: sample_id, signer_ref, start_frame, end_frame, primary_gloss, English meaning if needed, nonmanual marker, dominant_hand, occlusion flag, reviewer, review_date, qa_status.
- A note such as eyebrows raised during question can be important even if the gloss tokens are the same.
Practice
- Design an annotation form for one sign clip.
- Include at least one field for language meaning, one for timing, one for pose quality, and one for reviewer notes.
Checkpoint
Before moving on
You can explain why annotation is labor-intensive and language-sensitive.
Deaf-first note
Deaf-first note
When possible, include Deaf reviewers or fluent ASL users in annotation design, not only at the end.