Computer vision overview

Computer vision teaches machines to interpret images and video: what is there, where it is, and sometimes what will happen next.

Vision powers photo apps, industrial inspection, medical imaging assistance, robotics perception—and surveillance risks. Pair with Course 19/29 for misuse literacy.

Pipelines often: pixels → features (classical or learned) → task head (classify, detect boxes, segment masks, caption). Modern systems lean on CNNs and vision transformers.

Phone camera detecting a QR code or suggesting a crop—narrow vision tasks, not “understanding the whole world.”

Pick a photo and list: classify, detect, segment, caption. Which task matches your need?

⚠️ Myth: Vision models see like humans.

✓ Reality: They optimize training objectives; adversarial patches and dataset gaps reveal differences.

⚠️ Myth: Higher accuracy on a benchmark means safe deployment.

✓ Reality: Domain shift and demographic performance gaps still bite.

• Course 05 cnns-and-vision
• CS231n-style curricula (verify current offering)
• NIST AI RMF for risk framing: https://www.nist.gov/itl/ai-risk-management-framework ↗