Week 3: Training and Review
Day 18: Editing and Correcting Signs
Goal
Learn what correction can involve.
Learn
- Correction may mean trimming the clip, fixing labels, removing bad frames, re-running pose extraction, correcting keypoints, smoothing motion, or rejecting the sample.
- The goal is not to make the data look pretty. The goal is to make the record truthful, traceable, and useful for the model task.
- Every correction should leave a note so future reviewers know what changed.
Example
- Failure: the clip starts before the signer is ready. Correction: trim start frames and update frame_start.
- Failure: right hand keypoints jump to the face for ten frames. Correction: try re-extraction, then manual fix or reject if the sign meaning is affected.
- Failure: gloss says PLEASE but video is THANK-YOU. Correction: fix the label or remove the sample from training.
Practice
- Pick three failure types: bad trim, wrong gloss, missing fingertips.
- For each one, write the correction action and the metadata note you would save.
Checkpoint
Before moving on
You can match common data errors to correction steps.
Pipeline note
Pipeline note
Do not overwrite original source files. Keep raw data separate from corrected training records.