Review: Practical Details

Big failures often come from small details. A wrong resize. A colour order. A forgotten lens. Let's see if you'd catch them.

• Small details
• Big failures
• Would you catch them?

Recap: the details

Use area interpolation to shrink, and nearest for masks. Include compression in your training if you'll see it in production. Remember that OpenCV reads blue, green, red. Work out your receptive field. Budget every millisecond. And before you train at all, ask whether you can simply change the problem.

Question 1

Question 1. Which method suits resizing a label mask? A, Bicubic. B, Nearest neighbour. C, Area. D, Lanczos. Pause and choose. The answer is B. Averaging would invent classes.

Question 2

Question 2. What colour order does OpenCV use by default? A, RGB. B, BGR. C, HSV. D, CMYK. Pause and choose. The answer is B. That's caused many bugs.

Question 3

Question 3. What's the first thing to try for a hard defect? A, A bigger model. B, Better lighting. C, More epochs. D, A new optimiser. Pause and choose. The answer is B. Solve it at the camera.

Question 4

Question 4. Your network misses large defects. What should you check? A, The file format. B, The receptive field. C, The learning rate only. D, The colour order. Pause and choose. The answer is B. A neuron must be able to see the object.

Question 5

Question 5. Twelve apples a second gives what budget per apple? A, About 8 milliseconds. B, About 83 milliseconds. C, 1 second. D, 12 milliseconds. Pause and choose. The answer is B. Keep a margin too.

Question 6

Question 6. When should you use a general vision model rather than your own? A, Millisecond speed. B, Quick prototypes. C, High volume. D, Exact locations. Pause and choose. The answer is B. It's great for getting started.

Summary