All concepts

YOLO Object Detection

Single-pass CNN detector: grid cells predict boxes, NMS removes duplicates

Deep Learning · Intermediate · ~5 min

In plain English

Instead of scanning an image many times looking for objects, cut it into a grid and have every cell answer once: 'is something here, what, and where exactly?'

Why it's worth your time

It's the reason object detection runs in real time on a phone or a camera feed, and the standard answer for any 'find the things' problem.

If you remember three things

  • One forward pass predicts every box and class
  • Non-max suppression removes duplicate boxes for one object
  • Speed/accuracy is a model-size choice, not an architecture change

Overview

YOLO ('You Only Look Once') is a fast, single-stage object detector. Instead of the slow two-stage approach of proposing regions and then classifying them, a CNN backbone processes the whole image in one forward pass and outputs a grid whose cells each predict bounding boxes, an objectness score, and class probabilities. Because that yields many overlapping boxes, Non-Max Suppression keeps the best per object, giving real-time detection.

In an interview

YOLO is a real-time object detector that looks at the whole image once. A single CNN forward pass produces an S-by-S grid; each cell predicts a few bounding boxes with an objectness score and class probabilities. That gives thousands of overlapping boxes, so Non-Max Suppression keeps the highest-confidence box per object and drops the rest. Being one-stage — no separate region-proposal step — is what makes it fast enough for video, unlike two-stage detectors like R-CNN.

Production defaults

Confidence
0.25 for recall-first, 0.5 for precision-first. Set from your cost of a miss vs a false alarm
NMS IoU
0.45. Lower it when objects genuinely overlap and boxes are being merged wrongly
Data
fine-tune from COCO weights; a few hundred labelled images per class is often enough

What breaks

  • Misses small objects — Input resolution. Raise it, or tile the image and run detection per tile.
  • One object gets several boxes — NMS threshold too high, or classes are competing. Check class-agnostic vs per-class NMS.

Watch it explained

How YOLO works for object detection | Computer Vision — AI Sciences, 5:05

Related