Quiz: Deep Vision

This round covers deep vision, the kind of questions you'd meet in an interview. Take your time.

• Detection, segmentation, and more
• Interview level
• Take your time

Question 1

Question 1. What does IoU measure? A, Overlap between two boxes. B, Image size. C, Training speed. D, Colour match. Pause the video and choose. The answer is A. Intersection divided by union.

Question 2

Question 2. Why do detectors need NMS? A, To remove duplicate boxes. B, To add colour. C, To train faster. D, To resize images. Pause the video and choose. The answer is A. Several boxes often fire on a single object.

Question 3

Question 3. mAP is built from which two ideas? A, Precision and recall. B, Colour and shape. C, Speed and size. D, Noise and blur. Pause the video and choose. The answer is A. It's the area under the precision recall curve.

Question 4

Question 4. Why is pixel accuracy a poor metric for segmentation? A, Background dominates. B, It is slow. C, It needs a GPU. D, It ignores colour. Pause the video and choose. The answer is A. A model predicting all background can score highly.

Question 5

Question 5. You've got 300 images for a new task. What's your first step? A, Train a network from scratch. B, Transfer learning. C, Buy a GPU cluster. D, Collect a million more. Pause the video and choose. The answer is B. Pretrained features do most of the work.

Question 6

Question 6. What does a tracker add to a detector? A, Identity across frames. B, Colour. C, Resolution. D, Depth. Pause the video and choose. The answer is A. The same apple keeps the same ID.

Question 7

Question 7. What's special about CLIP? A, Images and text share one space. B, It only detects faces. C, It needs no images. D, It is a camera. Pause the video and choose. The answer is A. So you can classify and search using words.

Question 8

Question 8. For ten apples a second on a farm, which is best? A, A small trained model. B, A cloud LLM for each apple. C, A human. D, No model. Pause the video and choose. The answer is A. It's fast, cheap, and it works offline.

Question 9

Question 9. Your model drops in winter. What likely changed? A, The lighting, causing drift. B, The Python version. C, The file name. D, The labels only. Pause the video and choose. The answer is A. Image statistics shifted.

Question 10

Question 10. What's the risk with synthetic training data? A, The gap to reality. B, Too many labels. C, It is always free. D, It uses no compute. Pause the video and choose. The answer is A. Always test on real images.

Summary