Stop 10 · Computer VisionTopics · 4/9

Object Detection (YOLO, Overview)

In short

Classification says 'there's a dog'. Detection says 'dog here, bike there, person over there' with boxes around each. That's what powers traffic cameras, shelf scanners and self-checkout.

Locating and classifying multiple objects within an image with bounding boxes — YOLO's single-pass design made real-time detection practical.

The blob as a traffic cop points at a car, an auto and a cow, each outlined by a rectangle.

Think of it like this

The traffic cop's single glance

A traffic cop at a busy junction doesn't study each vehicle one by one. One glance and he knows: car on the left, auto in the middle, a cow blocking lane two. YOLO, short for You Only Look Once, works the same way: one pass, every object at once.

Watch it happen

How YOLO turns one image into boxes

Image

Backbone

Predictions

Threshold

NMS

1/5A 640x640 street photo goes in. One forward pass, no sliding windows.

The big ideas

What you need to know

Box plus label plus score

Each detection is a rectangle, a class like 'car', and a confidence from 0 to 1.

One-stage vs two-stage

YOLO predicts everything in one pass, so it's fast. Older two-stage models like Faster R-CNN propose regions first, then classify.

IoU measures overlap

Intersection over Union compares two boxes. NMS uses it to drop duplicates, and evaluation uses it to judge hits.

mAP is the score

Mean Average Precision summarises precision and recall across classes and IoU thresholds. Higher is better.

See it in code

Walk through the code

python

1# pip install ultralytics2from ultralytics import YOLO3 4model = YOLO("yolov8n.pt")            # small model pretrained on COCO5results = model("street.jpg", conf=0.25)6 7for box in results[0].boxes:8    cls_name = model.names[int(box.cls)]9    x1, y1, x2, y2 = box.xyxy[0].tolist()10    print(f"{cls_name} {float(box.conf):.2f} at {x1:.0f},{y1:.0f}")11 12results[0].save("street_boxes.jpg")

1/3Load the nano model and run it with a 0.25 confidence threshold. NMS happens inside.

Ultralytics runs a pretrained YOLO in a few lines. COCO covers 80 everyday classes like person, car and dog.

Save yourself the pain

Common mistakes

Oops

Judging a detector by how many boxes it draws.

Do this instead

Measure precision and recall at a fixed IoU, or mAP. Many low-confidence boxes usually means many false alarms.

Oops

Expecting a COCO model to find classes it was never trained on, like auto-rickshaws or specific products.

Do this instead

Fine-tune on a few hundred labelled images of your own classes.

Your mission

Count vehicles in a photo

~25 min

Quick check

Your detector draws three nearly identical boxes around one car. Which step is meant to fix this?

Go deeper

1Run a pretrained YOLO model on your own images or video.
2Learn the difference between classification, detection and segmentation tasks.
3Understand bounding box formats and non-max suppression at a conceptual level.