Day 15: Dataset Curation

Goal

Understand dataset curation as careful selection and organization.

Learn

• Curation means deciding what belongs in the dataset and what should be removed, repaired, or held for review.
• A curated dataset checks duplicates, label consistency, signer coverage, lighting variety, license status, consent, and class balance.
• More data is not automatically better. A smaller reviewed dataset can be more useful than a large folder of messy clips.

Example

• Keep: signer centered, both hands visible, correct gloss, consent present.
• Review: one hand briefly hidden, uncertain gloss, low pose confidence.
• Reject: no consent, face missing when facial grammar matters, wrong label, corrupted file.

Practice

1. Create keep/review/reject rules for a small sign clip dataset.
2. Apply those rules to three pretend samples.

Checkpoint

Pipeline note