AI Hints
Home›Computer Vision Tutorial›Review: Foundations
Computer Vision · Lesson 98 of 111

Review: Geometry and Features

How vision measures the world.

⏱ 2 min read▶ Video at 2:36:24🎓 Free
Watch this lesson in the Computer Vision courseWatch on YouTube ↗Full Computer Vision course

Geometry is what turns a picture into a measurement. It's often the part of vision that gets skipped, and it's where you can stand out. Let's check what you've kept.

Review: Geometry and Features
From the video: Review: Geometry and Features
  • Pictures become measurements
  • Often skipped
  • Where you can stand out

Recap: geometry

Keypoints are distinctive spots that can be matched between images. The ratio test filters out poor matches, and RANSAC finds the transform that most matches agree on. A homography maps one flat plane onto another, from four points. Calibration measures a lens, from a checkerboard. And two cameras give you depth, by triangulation.

Keypointsmatch between images
RANSACignores bad matches
Homographyfour points, flat planes
Calibrationcheckerboards

Question 1

Question 1. How many point pairs define a homography? A, Two. B, Three. C, Four. D, Nine. Pause and choose. The answer is C. Four pairs give the eight numbers that you need.

Quick check
How many point pairs define a homography?

Question 2

Question 2. What does RANSAC assume? A, All data is good. B, Much of the data may be wrong. C, Images are flat. D, Data is sorted. Pause and choose. The answer is B. It looks for the largest group that agrees.

Quick check
What does RANSAC assume?

Question 3

Question 3. A homography only works perfectly for what? A, Flat surfaces, or pure rotation. B, Any scene. C, Moving objects. D, Black and white images. Pause and choose. The answer is A. 3D scenes break it.

Quick check
A homography only works perfectly for what?

Question 4

Question 4. What does a low reprojection error tell you? A, The photos are sharp. B, The camera model explains the data well. C, The lens is cheap. D, The image is small. Pause and choose. The answer is B. Under half a pixel is good.

Quick check
What does a low reprojection error tell you?

Question 5

Question 5. What should you do after calibrating? A, Change the zoom. B, Lock the focus and zoom. C, Move the camera. D, Add noise. Pause and choose. The answer is B. Changing them means calibrating again.

Quick check
What should you do after calibrating?

Question 6

Question 6. What gives depth from two cameras? A, Disparity between views. B, Colour. C, Brightness. D, Compression. Pause and choose. The answer is A. Closer objects shift more.

Quick check
What gives depth from two cameras?

Summary

Next, deep vision.

Key takeaways
  • RANSAC handles bad matches
  • Homographies need flat surfaces
  • Calibration needs locked lenses
Next up: Review: Deep Vision — classifiers, detectors, and masks