Day 4: The Basic Pipeline

Goal

Learn the full pipeline at a high level.

Learn

• A common pipeline is: capture video, extract frames, detect pose/keypoints, save structured data, add labels or gloss, clean the dataset, train the model, validate it, and manually inspect the outputs.
• Each step can introduce errors. Bad capture affects pose. Bad pose affects training. Bad labels teach the model the wrong target.
• The pipeline is not finished when a model runs. It is finished only after review shows that the output is useful, honest, and safe for the intended purpose.

Example

• One sample may start as clip_0042.mp4. Frames are extracted, DWPose or another pose tool creates keypoints, the annotator labels the clip as THANK-YOU, and the final training record points to clip_0042.npz.
• If the signer moves out of frame, the sample may be marked review or reject instead of going into training.

Practice

1. Draw this flow: capture -> frames -> pose/keypoints -> gloss/labels -> NPZ -> train -> validate -> human review.
2. Add one possible failure under each step.

Checkpoint

Pipeline note