culane_metric: let a prediction claim a free lane instead of dropping it

[egeboy35]

The module docstring promises a "greedily one-to-one match". _match offers each prediction only its argmax ground-truth lane:

for i in np.argsort(-iou.max(axis=1)):
    j = int(iou[i].argmax())
    if j not in used_g and iou[i, j] >= iou_thresh:

So when that lane is already taken the prediction is discarded — including where the same IoU matrix says it clears the threshold against a lane nothing has claimed. Greedy one-to-one takes that pair.

Measured

Found by sweeping lane spacing against prediction offset at the class's own defaults (320×800, width 15 px). Ground truth at x=300 and x=306, predictions at x=300 and x=301:

IoU     g0       g1
  p0   1.000*   0.478
  p1   0.889*   0.545*        * clears the 0.5 threshold

p0 takes g0. p1's argmax is also g0, now used, so p1 is dropped — while IoU(p1, g1) = 0.545 and g1 is free.

_match              tp=1  fp=1  fn=1
greedy one-to-one   tp=2  fp=0  fn=0

The missed pair is charged twice: once as a false positive, once as a false negative. Across lane gaps 4–17 px and prediction offsets in 0.5 px steps, 29 of 294 combinations under-count.

Over 400 random four-lane frames through the public CULaneF1 class, where a quarter of the predictions are pulled toward a neighbouring lane:

_match              tp=1218  fp=382  fn=382   F1=0.7612
greedy one-to-one   tp=1236  fp=364  fn=364   F1=0.7725

Both matchers see the same IoU matrices, so the difference is the matching rule alone — not the geometry, the threshold or the model. F1 is understated by 0.0113 on that sample, and the direction is fixed: the previous rule can never count more.

The change

Walk the matrix best-pair-first and skip a pair only when the prediction or the lane is already used — the docstring's rule applied to the whole matrix rather than to one column per row. It stays pure numpy, and the loop breaks at the first sub-threshold pair because the order is sorted.

What I am not claiming

Greedy is not always the maximum matching. It can commit a high pair that blocks two lower ones. On the 40-frame sample in the tests, greedy takes 129 where an exhaustive search takes 130.

Closing that gap means Hungarian assignment, which would pull scipy into a module whose first line advertises "pure numpy (no cv2, no external eval binary)", and would depart from the rule the docstring states. If you would rather have the optimum I am happy to send that instead — it is a small change on top of this one.

Tests

Adds ngperception/lane/tests/test_culane_metric.py — 12 tests, no dataset, no model, no GPU.

Against this branch: 12 passed. Against the file as it stands on main: 3 failed, 9 passed:

FAILED test_a_prediction_is_not_dropped_while_a_lane_it_matches_is_free
FAILED test_the_sweep_that_found_it_reports_no_under_counting
FAILED test_the_class_agrees_with_an_independent_greedy_matcher   (124 vs 129)

The 9 that pass either way are the properties this must not break: one lane per prediction, one prediction per lane, counts adding up, the threshold respected, the empty cases, and a perfect prediction scoring 1.0.

pytest DeepDataMiningLearning/ngperception/lane/tests