Design a cashierless retail checkout vision system.

WHAT THIS TESTS The interviewer evaluates whether you can architect a complete multi-component vision system, reason about deployment, and plan for the messy real world rather than naming one model.

A GOOD ANSWER COVERS Camera layout: overhead cameras for tracking people across the floor plus shelf-facing cameras for fine product interaction, arranged for overlapping coverage to reduce blind spots. Core vision tasks: person detection and multi-object tracking, with person re-identification to maintain identity across cameras and time; product detection and classification on shelves; action or hand-interaction recognition to decide whether an item was taken or returned; and association logic linking each item event to the correct shopper's virtual cart. Sensor fusion: combine vision with shelf weight sensors or RFID to disambiguate hard cases. Deployment: run latency-sensitive detection and tracking on edge devices near the cameras for real-time response and bandwidth savings, while heavier re-ID matching and analytics can run in the cloud; balance latency, bandwidth, cost, and privacy. Failure modes: occlusion and crowding, identity swaps during close contact, look-alike products, and ambiguous take-or-return; mitigate with multi-view redundancy, sensor fusion, confidence thresholds, and routing low-confidence events to human review or receipt correction.

COMMON WRONG ANSWERS Proposing one detector and calling it done. Ignoring person tracking and item-to-person association. No fusion, no failure handling, no edge-versus-cloud reasoning.

LIKELY FOLLOW-UPS Handling re-ID under similar clothing, the privacy and data-retention implications, latency budgets, and how to bootstrap training data.

ONE CONCRETE EXAMPLE In a Just Walk Out style store, overhead cameras track each shopper, shelf cameras plus weight sensors detect a can being picked up, action recognition confirms a take rather than a return, association adds it to that person's cart, and an ambiguous double-grab is flagged for review so the final receipt stays correct.