Module 05 Activity

Scenario

The dataset is where most vision projects are actually won or lost.

What you build

A labelled dataset with edge-case rulings, a grouped split, a near-duplicate check and a measured label agreement rate.

Steps

1. Write the label schema with one worked example per class and rulings on three genuinely borderline cases. Do this before labelling.
2. Report the class distribution and compute the majority-class baseline.
3. Split by the underlying object or scene, not by file. Prove no group appears on both sides.
4. Check every augmentation you plan to use against the label it must preserve. Remove the ones that corrupt it.
5. Have a second person re-label 60 images and report the agreement rate.
6. If agreement is below 85%, refine the rulings and re-audit.

Evidence to hand in

• The schema with worked examples and three edge-case rulings.
• The class distribution and majority baseline.
• The grouped split with proof of no overlap.
• The augmentation audit with anything removed.
• The agreement rate, and the re-audit if one was needed.

Review checklist

• Edge-case rulings were written before labelling.
• The split is by object or scene, not by file.
• Every augmentation was checked against the specific label.
• Agreement was measured by a second person on a real sample.
• A low agreement rate led to refinement, not to proceeding.

Mark complete