Anchor Box Matching for Object Detection

Feature Map

Grid resolution 6 × 6

Anchor base scale 1.00×

IoU Assignment Rule

Positive threshold 0.50

Negative threshold 0.30

Live Readouts

Anchors

Positive

Best IoU

0.00

Positive (matched) anchor

Negative anchor

Ignored (mid-IoU) anchor

Ground-truth box

How it works

A one-stage detector (SSD, RetinaNet, the original Faster R-CNN RPN) never predicts a box from scratch. It tiles a dense set of reference rectangles — anchor boxes — over every cell of a convolutional feature map, at several scales and aspect ratios, then asks each anchor two questions: "is an object here?" and "how should I nudge my edges to fit it?"

IoU(A, G) = area(A ∩ G) / area(A ∪ G)

anchor size  = cellSize × scale × √ar  (width)
             = cellSize × scale / √ar  (height)
scale ∈ {0.7, 1.0, 1.3} × base scale
ar    ∈ {1:1, 2:1, 1:2}

Every anchor's IoU with the ground-truth box decides its training label:

Positive — IoU ≥ positive threshold, or it is the single anchor with the highest IoU overall (so every object gets at least one match, even a sliver of one).
Negative — IoU below the negative threshold; trained to predict "no object."
Ignored — IoU between the two thresholds; excluded from the loss so the network isn't punished for a genuinely ambiguous overlap.

This is exactly why detectors need N anchors per cell instead of one: a tall anchor and a wide anchor at the same location cover very different real objects, and only the shapes close to the true box light up.

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.