Module 08 Activity

Scenario

Detection adds boxes, and boxes add three new ways to be wrong.

What you build

A detection evaluation with an explicit box format, a stated IoU threshold, working suppression, and a small-object assessment.

Steps

1. State your box format explicitly and convert at the boundary. Assert the format once, loudly.
2. Compute IoU for four predictions against one ground truth and show how the hit/miss verdict changes between 0.5 and 0.75.
3. Implement non-maximum suppression and report detections before and after on a crowded image.
4. Report precision and recall at four confidence thresholds, and choose one from the cost of each error.
5. Measure your smallest target object in pixels after your resize.
6. If it is under a few pixels, design a tiling approach rather than a larger model.

Evidence to hand in

• The box format and the boundary assertion.
• IoU at two thresholds with the changed verdicts.
• Detection counts before and after suppression on a crowded image.
• Precision and recall at four thresholds with your chosen operating point.
• The smallest object's post-resize pixel size and your response.

Review checklist

• The box format is asserted, not assumed.
• The IoU threshold is reported with every metric.
• Suppression was tuned on crowded images, not well-separated ones.
• The threshold follows from error costs.
• Small-object handling is tiling, not a bigger backbone.

Mark complete