6.15 Object detection and segmentation
You can go beyond 'what is in the image' to 'where'.
Before:05. Classical Machine LearningUnlocks:07. Natural Language Processing11. The ECE Bridge — Signals, Edge & Embedded AI
Detection and segmentation move from what to where: boxes with the IoU, NMS and mAP machinery, masks per pixel, two-stage against one-stage designs — and promptable models like SAM shifting what needs training at all. It sits in the advanced vision run because it stacks on classification. The underestimated part is evaluation: mAP has more moving parts than most models, and reading it wrongly makes bad detectors look good.
Work through these
Bounding boxes, IoU, NMS, mAP
The vocabulary of detection: boxes around objects, the overlap measure, the procedure that removes duplicate boxes, and the summary score. You cannot read a detection paper without these four.
Two-stage (Faster R-CNN) vs. one-stage (YOLO, RetinaNet)
Detectors that propose regions and then classify them, against detectors that predict everything in one pass. The trade is accuracy against speed, and it has narrowed over time.
Semantic vs. instance segmentation; U-Net, Mask R-CNN
Labelling every pixel with a class, against separating individual objects, and the architectures associated with each. The distinction decides which is the right task for your problem.
Segment Anything and promptable segmentation
Models that segment whatever you point at, without being trained on your categories. This changed what a segmentation project has to start from.
Sign in to keep your progress.
Free resources
We haven't checked most of these for screen reader use yet.
Links last checked 29 Aug 2026.
Stuck here?
Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.
Checking your session…
Topics shown in module order.