Chapter 2: Deep learning
Detection metrics: IoU, NMS and mAP
A detector does not give one right-or-wrong answer: it draws many boxes, and each box has a score. You will be able to measure how well two boxes overlap, delete duplicate boxes, and turn a ranked list of boxes into one score. Interviewers often ask you to code the first two and to explain the score.
What you will learn
- Compute IoU between two boxes
- Implement greedy non-maximum suppression
- Compute average precision for one class from ranked detections
Key terms
- bounding box
- a rectangle around an object, written [x1, y1, x2, y2]
- ground truth
- the true box that a person drew around a real object
- IoU
- a score from 0 to 1 for how much two boxes overlap
- NMS
- a step that deletes duplicate boxes for the same object
- mAP
- one score for a whole detector, found from ranked boxes
Prerequisites
Take Convolutions, residual blocks and conv arithmetic first. It comes right before this lesson in the course.
Where this lesson picks up: You saw how conv layers turn a photo into feature maps, and you worked out their output shapes. A detector reads those maps and draws boxes with scores, so how do we decide if the boxes are good?
What comes next
Next, you will ask whether the feature maps should come from convolutions or from a newer design, the vision transformer.
Next lesson: CNNs vs ViTs and transfer learning
More lessons in this part
Take the lesson
Short cards, one idea at a time, with quick checks as you go.