Review: Deep Vision

Deep learning does the heavy lifting in modern vision. This round covers the core ideas, and the scores that tell you if it's working.

• The heavy lifting
• Core ideas
• The scores that matter

Recap: deep vision

Convolutions share weights, and find local patterns. Transfer learning reuses a trained network, and it's your default start. Detectors output boxes, with confidence, and non maximum suppression removes duplicates. Average precision summarises the detector. Segmentation labels every pixel, and U-Net uses skip connections to keep fine detail.

Question 1

Question 1. Why are convolutions efficient? A, They share weights across the image. B, They use no memory. C, They skip pixels. D, They need no training. Pause and choose. The answer is A. The same filter runs everywhere.

Question 2

Question 2. What does non-maximum suppression do? A, Removes overlapping duplicates. B, Increases contrast. C, Trains a model. D, Resizes images. Pause and choose. The answer is A. It keeps the highest scoring box.

Question 3

Question 3. What does mAP@50 require for a correct box? A, 50 percent confidence. B, Overlap of at least 0.5. C, 50 objects. D, 50 epochs. Pause and choose. The answer is B. It's about overlap with the truth.

Question 4

Question 4. Why not use pixel accuracy for tiny objects? A, It is slow. B, Predicting nothing scores very high. C, It needs labels. D, It is random. Pause and choose. The answer is B. Use Dice or IoU.

Question 5

Question 5. What do U-Net's skip connections carry? A, Labels. B, Fine spatial detail. C, Errors. D, Colour tables. Pause and choose. The answer is B. They bring detail across to the decoder.

Question 6

Question 6. What's the first step for a small dataset? A, Build a huge model. B, Transfer learning. C, Train from scratch. D, Buy GPUs. Pause and choose. The answer is B. Pretrained features save a huge amount of data.

Summary