[AryanRG]

Summary

Handle zero-width quadrilaterals in horizontal crop extraction without raising IndexError. Return the existing empty-crop placeholder so predictor filtering keeps crops, boxes and objectness scores aligned. The oriented crop path is unchanged.

The regression cases are folded into the existing geometry and OCR/KIE score-alignment tests, covering both straight and rotated boxes.

Boxes with NaN x coordinates also produce an empty crop; broader non-finite polygon validation is outside this patch.

Validation

Focused geometry, OCR/KIE alignment and adjacent builder/utility tests pass locally, along with pre-commit, Ruff and mypy. The consolidated regressions reproduce the original failure on upstream code. The full local suite was not rerun for this test-only revision.

[felixdittrich92]

Hi @AryanRG 👋

Thanks for the fix.

Tests: please merge them into the existing ones

1. test_extract_rcrops_zero_width → fold into test_extract_rcrops

The new test runs 16 parametrized cases (4 boxes × 2 paths × 2 image types), and much of it checks things this PR doesn't touch:

• half the cases run the oriented path (assume_horizontal=False), which is unchanged;
• the grayscale/float32 image, the input-not-modified checks and the comparison against single-box crops all cover existing behaviour, not the fix.

test_extract_rcrops is already parametrized over assume_horizontal, so a few lines there cover the regression on both paths. Add this just before the # No box check:

    # Zero-width boxes yield an empty crop, without shifting the other crops
    collapsed_boxes = np.array(
        [[[x, 0.25], [x, 0.25], [x, 0.75], [x, 0.75]] for x in (0.0, 0.5, 1.0)] + [[[0.5, 0.5]] * 4],
        dtype=np.float32,
    )
    croped_imgs = geometry.extract_rcrops(
        doc_img, np.concatenate([rel_boxes, collapsed_boxes]), assume_horizontal=assume_horizontal
    )
    assert len(croped_imgs) == num_crops + len(collapsed_boxes)
    assert all(crop.size > 0 for crop in croped_imgs[:num_crops])
    assert all(crop.shape == (0, 0, 3) and crop.dtype == doc_img.dtype for crop in croped_imgs[num_crops:])

Then remove test_extract_rcrops_zero_width.

2. test_predictors_drop_zero_width_quadrilaterals → parametrize test_predictors_keep_objectness_scores_aligned

The new test is about 70 lines and largely repeats test_predictors_keep_objectness_scores_aligned: the same fixed detection predictor, the same dummy recognition predictor, the same OCR/KIE parametrization and the same [0.9, 0.8] objectness assertions. The only new thing is the rotated-box input. Adding an assume_straight_pages parameter to the existing test covers it.

Some parts of the new test can be dropped:

• The "all-collapsed" cases only exercise the generic empty-crop filtering in _prepare_crops, which this PR doesn't change.
• The sort by minimum x isn't needed: words come back in box order and word.geometry[0][0] works for polygons too.
• The reco call counting and the crop.size > 0 check inside the recognizer are already implied by the word and score assertions.
• The type(result) is ... check is already covered by test_predictors_on_empty_batch.
• detect_orientation=False, straighten_pages=False, resolve_lines=False, resolve_blocks=False are either defaults or don't affect the result.

Suggested replacement for test_predictors_keep_objectness_scores_aligned:

@pytest.mark.parametrize("assume_straight_pages", [True, False])
@pytest.mark.parametrize("predictor_cls", [OCRPredictor, KIEPredictor])
def test_predictors_keep_objectness_scores_aligned(predictor_cls, assume_straight_pages):
    # The middle box lies on the right border of the page and has no width: its crop is empty,
    # so it is dropped before recognition
    if assume_straight_pages:
        boxes = np.array(
            [[0.1, 0.1, 0.3, 0.2, 0.9], [1.0, 0.5, 1.0, 0.6, 0.1], [0.5, 0.1, 0.7, 0.2, 0.8]],
            dtype=np.float32,
        )
    else:
        # Polygons with the objectness score in the last row
        boxes = np.array(
            [
                [[0.1, 0.1], [0.3, 0.1], [0.3, 0.2], [0.1, 0.2], [0.9, 0.9]],
                [[1.0, 0.5], [1.0, 0.5], [1.0, 0.6], [1.0, 0.6], [0.1, 0.1]],
                [[0.5, 0.1], [0.7, 0.1], [0.7, 0.2], [0.5, 0.2], [0.8, 0.8]],
            ],
            dtype=np.float32,
        )

    class _FixedDetPredictor(nn.Module):
        def forward(self, pages, **kwargs):
            return [{CLASS_NAME: boxes.copy()} for _ in pages]

    class _DummyRecoPredictor(nn.Module):
        def forward(self, crops, **kwargs):
            return [(f"word{idx}", 1.0) for idx in range(len(crops))]

    # Without page orientation, rotated boxes are cropped with `assume_horizontal=True`
    predictor = predictor_cls(
        _FixedDetPredictor(),
        _DummyRecoPredictor(),
        assume_straight_pages=assume_straight_pages,
        disable_page_orientation=True,
        disable_crop_orientation=True,
    )
    page = predictor([np.full((100, 200, 3), 255, dtype=np.uint8)]).pages[0]
    if predictor_cls is KIEPredictor:
        words = page.predictions[CLASS_NAME]
    else:
        words = [word for block in page.blocks for line in block.lines for word in line.words]

    # Each remaining word keeps the objectness score of its own box
    assert [word.geometry[0][0] for word in words] == pytest.approx([0.1, 0.5])
    assert [word.objectness_score for word in words] == pytest.approx([0.9, 0.8])

disable_crop_orientation=True also keeps the test from loading the pretrained crop orientation model. Then remove test_predictors_drop_zero_width_quadrilaterals.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.61%. Comparing base (11941d2) to head (126e127).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2173   +/-   ##
=======================================
  Coverage   97.61%   97.61%           
=======================================
  Files         170      170           
  Lines       10740    10744    +4     
=======================================
+ Hits        10484    10488    +4     
  Misses        256      256           

Flag	Coverage Δ
unittests	97.61% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[AryanRG]

Thanks @felixdittrich92, folded both regressions into the existing tests and removed the duplicates. Tests and checks pass locally.

[felixdittrich92]

Thanks 👍