fix(iou): clamp box extents so a malformed box cannot earn a GIoU bonus

[MahathirMohammadShuvo]

The defect

_compute_iou_and_enclosing computes box area as (x2 - x1) * (y2 - y1) with no clamp — three lines after the intersection is correctly clamped with np.maximum(..., 0):

intersection = np.maximum(inter_x2 - inter_x1, 0) * np.maximum(inter_y2 - inter_y1, 0)   # clamped

area_1 = (boxes_1[:, 2] - boxes_1[:, 0]) * (boxes_1[:, 3] - boxes_1[:, 1])               # not clamped

For a box inverted on both axes the two negative extents multiply to a positive area. That inflates union past enclosing_area, so GIoU's (enclosing - union) / enclosing penalty term goes negative — a bonus instead of a penalty.

Reachable through a shipped tracker

SORTTracker's default XYXYStateEstimator has an unconstrained velocity (clamp_velocity is a documented no-op — "XYXY-style velocity is unconstrained"), and sort/tracklet.py has no downstream box clamp: state_to_bbox() returns kf.state[:4] straight into self.iou.compute.

Four shrinking detections, then six unmatched frames:

tracker = SORTTracker(iou=GIoU())
# ... tracklet bbox -> [274.3, 274.3, 125.7, 125.7]     x2 < x1 and y2 < y1

	GIoU(ghost box, real detection)	GIoU(well-formed near-perfect match)
before	+2.2080	+0.6400
after	+0.0000	+0.6400

The ghost outranks a genuine match, at a score outside GIoU's documented [-1, 1]. Over 3,000 random malformed pairs the pre-fix maximum was +29.54.

Scope

• GIoU only. IoU/BIoU route through sv.box_iou_batch and never enter this helper. DIoU/CIoU are byte-identical before and after, because a malformed axis already forces intersection = 0 and they consume only iou and enclosing_diagonal_sq.
• Well-formed results are bit-identical — np.maximum(x, 0) returns x unchanged for x > 0 — so no benchmark run is needed.
• Not fixed here: DIoU/CIoU remain unbounded below on malformed input, because the enclosing box is built from min/max of the corners, so an inverted box makes C stop enclosing the pair and the distance normaliser collapses ([500,500,0,0] vs [0,0,10,10] → DIoU -600.25). That is a separate defect in a different term. The docstrings now say so rather than implying it is covered — happy to fold it into this PR if you would prefer.
• I first tried guarding the producing estimators (XYXYStateEstimator / XCYCWHStateEstimator both have no-op clamp_velocity) and abandoned it: freezing extent velocity changes tracking output on the default frame_step=1.0 path for BoT-SORT/McByte/C-BIoU, and can freeze the velocity that would recover an inverted box. Fixing the metric covers every producer at once.

Tests

TestDegenerateInputs.test_inverted_coords_gives_zero_or_negative_similarity promised its name but asserted only shape and finiteness, so it passed at GIoU +1.0. It is now split into an exact-value test (all three inversion patterns must equal a well-formed zero-area box) and a sign test across IoU/GIoU/DIoU/CIoU/BIoU × 3 patterns, plus a union <= enclosing_area invariant.

5 assertions fail on develop without the source change, including assert 4000.0 <= 3200.0.

ruff check clean · ruff format clean · mypy clean · iou.py doctests 6/6 · tests/utils + tests/motion 165 passed · tests/core fast subset 607 passed.

[CLAassistant]

All committers have signed the CLA.