Object detection basics

Object detection finds *what* objects are present and *where* (usually bounding boxes). It is more than whole-image classification.

Retail analytics, accessibility tools, industrial QA, and many robotics stacks start with detection—along with serious misuse potential if aimed at people without consent.

Models propose regions or predict boxes densely across the image, score classes, then suppress duplicates (non-max suppression family). Training needs labeled boxes—expensive and error-prone.

A camera app drawing rectangles around faces for focus—detection + a policy choice about whether that is OK.

On a cluttered desk photo, count objects you’d need boxes for to build an inventory app. Labeling cost becomes obvious.

⚠️ Myth: A box means the model understands the object.

✓ Reality: It means localization + class score under training definitions.

⚠️ Myth: 99% accuracy on one dataset transfers everywhere.

✓ Reality: Camera angle, geography, and lighting shift performance.

• Course 12 overview; Course 05 CNNs
• Detection model cards / papers for the architecture you teach (YOLO/Faster R-CNN families—cite specifically)
• Course 19 bias chapter for demographic failures