AI Hints
Home›Computer Vision Tutorial›Myths and Design Interviews
Computer Vision · Lesson 103 of 111

Vision System Design

The interview question that matters.

⏱ 2 min read▶ Video at 2:52:36🎓 Free
Watch this lesson in the Computer Vision courseWatch on YouTube ↗Full Computer Vision course

A senior interview won't ask you what a convolution is. They'll say, design a system that counts people entering a shop. There's no single right answer. They want to see how you think. Here's a framework, with a full worked answer.

Vision System Design
From the video: Vision System Design
  • Design a people counter
  • No single right answer
  • How you think matters

The framework

Start with questions. What's the goal, how accurate must it be, and what can we change? Then think about the camera and the scene. Then the simplest possible solution. Then the model and the pipeline. Then how you'd measure it. And finally, what could go wrong in the real world.

  1. Clarifygoal and limits
  2. Scenecamera and light
  3. Simple firstbaseline
  4. Pipelinemodel and speed
  5. Failureswhat breaks?
The framework
From the video: The framework

A worked answer

First, ask why. Counting for staffing can tolerate five percent error, and no identities are needed. So mount the camera overhead, pointing straight down. That removes faces, and solves privacy, and stops people from hiding behind each other. Use a small detector, on an edge device. Track the heads, and count line crossings, with each track counted once. Store only counts, never video.

Goalstaffing, 5% error
Cameraoverhead, no faces
Modelsmall detector, edge device
Countingline crossing, once per track
Privacystore counts only

Measuring and failing

To measure it, count by hand for an hour each day, for two weeks, and compare the numbers. Now list the failures. Groups walking together. Pushchairs. Staff going in and out all day. Sunlight at closing time. Each gets a test. And a daily comparison with till receipts will warn you if the camera gets moved. That's the kind of thinking that gets people hired.

Measurehand counts daily
Failure casesgroups, pushchairs, staff
Lightinglow sun
Monitorcompare with tills

Summary

Clarify, simplify, and plan for failure. Next, another worked example, from a different world.

Key takeaways
  • Ask why first
  • Change the camera, solve the problem
  • Name the failures
Next up: Designing Visual Search — a second worked answer