Day 47: Anchor-free detection & how YOLO predicts boxes
How YOLO turns pixels into boxes
YOLO ('You Only Look Once') divides the image into a grid and, at each cell, predicts boxes directly: a center offset, a width and height, an objectness score (is there an object here?), and class probabilities. Older YOLO used anchors — predefined box shapes the model nudges — but modern anchor-free versions predict box geometry directly from each location, simpler and with fewer hyperparameters to tune.
This builds on Stage 1 directly
YOLO's backbone is a CNN — the convolutions, feature hierarchy, and receptive field you learned Days 34–36. The detection 'head' is just extra convolutional layers that output box coordinates and class scores instead of a single classification. You're not learning a new field; you're adding an output shape to what you already understand.
Because many grid locations near an object all fire, YOLO produces *lots* of overlapping candidate boxes for a single object. Turning that redundant set into one clean box per object is tomorrow's topic — non-max suppression.
Key terms
- Grid prediction
- Dividing an image into cells and predicting objects from each cell's features.
- Anchor box
- A predefined box shape a detector adjusts to fit objects; anchor-free methods skip these and predict geometry directly.
- Objectness score
- A predicted probability that a given location actually contains an object.
What does an anchor-free YOLO model predict at each grid location?
