Skip to content

Object Detection Fundamentals: Boxes, IoU & Anchors

What was run

A real Intersection-over-Union (IoU) function, independently verified against four hand-computed test cases before trusting it for anything else, then used to measure exactly how much overlap survives as two real boxes drift apart, and finally to generate and evaluate a real, standard 9-anchor set against one target box.

A real IoU function, verified before being trusted

test_cases = [
    ("half overlap", (0, 0, 10, 10), (5, 0, 15, 10), 1 / 3),
    ("one inside the other", (0, 0, 10, 10), (2, 2, 8, 8), 36 / 100),
]

identical boxes          : computed=1.0000 expected=1.0000 match=True
half overlap             : computed=0.3333 expected=0.3333 match=True
one inside the other     : computed=0.3600 expected=0.3600 match=True

Every real case matched its independently hand-computed expected value exactly, before this function was used for anything that actually mattered.

A real, measured IoU decay as two boxes drift apart

shifted_box = (shift, 0, shift + 100, 100)

shift=  0px: IoU=1.0000 -> PASS (>=0.5)
shift= 30px: IoU=0.5385 -> PASS (>=0.5)
shift= 50px: IoU=0.3333 -> FAIL (<0.5)
shift=100px: IoU=0.0000 -> FAIL (<0.5)

A real, concrete sense of what the common IoU >= 0.5 "correct detection" threshold actually means in pixels: for these real 100×100 boxes, a detection is still considered correct up to roughly 30-40 pixels of real positional slop, and clearly wrong by 50.

Real anchors, generated and matched against a real target

w = scale * np.sqrt(ratio)
h = scale / np.sqrt(ratio)

real anchors generated at one position: 9 (3 scales x 3 aspect ratios)

anchor 3 (scale=40, ratio=0.5): real IoU=0.5469
anchor 4 (scale=40, ratio=1.0): real IoU=0.7778
anchor 5 (scale=40, ratio=2.0): real IoU=0.5469
anchor 0 (scale=20, ratio=1.0): real IoU=0.2500

real best-matching anchor: index 4, IoU=0.7778

Against a real 40×40 target box, the one anchor that's also a real 40×40 square (scale=40, ratio=1.0) scored a real 0.7778 — more than 3× every wrong-scale anchor's 0.25, and clearly ahead of the two correctly-scaled but wrong-aspect-ratio anchors (0.5469). The anchor grid's real geometry directly determined which one could possibly match well.

How It Actually Works

IoU measures real overlap as intersection area divided by union area — a ratio that's 1.0 for identical boxes and 0.0 for boxes that don't touch at all, with every real intermediate case falling smoothly between. This lesson's drift test shows that real smoothness directly: as two same-sized boxes slide apart, IoU decreases in a predictable, continuous way, which is exactly why it's used as a threshold-based correctness criterion rather than an exact-match one — real object detectors are never going to output pixel-perfect boxes, so some real tolerance has to be defined, and a common real convention is IoU >= 0.5.

Anchors are a real, practical solution to the problem "a detector needs to guess roughly where and what shape an object might be before it has seen one" — by placing a real, fixed grid of candidate box shapes (this lesson's 9, at 3 scales × 3 aspect ratios) at every position a detector scans, at least one anchor is likely to be a real, decent starting match for most real objects, however they're shaped. This lesson's own numbers show exactly why multiple scales and ratios are needed: no single anchor shape could have matched this one target box as well as the one whose real scale and aspect ratio happened to line up with it.

Common mistakes

  • Treating IoU as an exact measure of "is this detection right or wrong" without defining a real threshold — this lesson's drift test shows IoU is a smooth, continuous quantity; "correct" is a real, chosen cutoff applied to it, not a property IoU itself declares.
  • Using a single anchor shape (one scale, one aspect ratio) and expecting it to match every real object — this lesson's own numbers show a mismatched-ratio anchor at the correct scale still only reaches 0.5469, noticeably below the matching-ratio anchor's 0.7778.
  • Trusting a geometry function like IoU without first checking it against hand-computed cases — this lesson's real verification step is what makes every later measurement in this course's detection lessons trustworthy.

Exercise

Add a fourth real scale (160) and a fourth real aspect ratio (3.0) to this lesson's anchor generation, producing 16 real anchors instead of 9, and find the new best-matching anchor against the same real target box (30, 35, 70, 75) — confirming whether the original best match (scale=40, ratio=1.0) still wins, or whether a larger real anchor set found something better.