Anchor Box IoU Field & Multi-Object Assignment (2D)

Feature Map

Grid resolution 6 × 6

Anchor base scale 1.00×

IoU Assignment Rule

Positive threshold 0.50

Negative threshold 0.30

Ground-truth objects 3

Live Readouts

Anchors

Positive

Ignored

Objects matched (recall)

–

Mean best IoU / object

–

Positive (matched) anchor

Negative anchor

Ignored (mid-IoU) anchor

Ground-truth boxes

How this 2D model differs from the 3D version

The 3D scene tiles one dense anchor grid over a single ground-truth box and lifts positive anchors into little green towers. This 2D companion computes the same underlying quantity — Intersection-over-Union between axis-aligned rectangles — independently and from scratch, but for the case real detectors actually train on: several ground-truth objects on the same image at once.

IoU(A, G) = area(A ∩ G) / area(A ∪ G)

For anchor A, best match over ALL objects:
  bestIoU(A) = max_g IoU(A, G_g)
  bestObj(A) = argmax_g IoU(A, G_g)

For each object G_g, per-object guarantee:
  argmax_A IoU(A, G_g)  is ALWAYS positive,
  even if that max is below the positive threshold —
  otherwise a small or sliver-thin object gets zero
  training signal and the detector never learns it exists.

Every anchor is labeled from its best-matching object only:

Positive — bestIoU(A) ≥ positive threshold, or A is some object's individual best-match anchor.
Negative — bestIoU(A) below the negative threshold against every object.
Ignored — bestIoU(A) between the two thresholds; excluded from the loss.

The strip beneath the grid is a live histogram of bestIoU(A) across every anchor — a distribution the 3D view never plots. Watch it shift right as objects grow larger or anchor scale is tuned to match them, and watch the "objects matched" readout stay pinned at 100% however tiny an object gets: that is the per-object guarantee working, verified numerically to hold for 1–4 simultaneous objects at any grid size.

Multi-object IoU field: each anchor is colored by its best match across all ground-truth boxes; bottom strip is the live bestIoU histogram.