OCR: boxes within a row are not sorted left-to-right

[msluszniak]

Description

useOCR().forward(image) returns detected text boxes sorted top-to-bottom, which is correct for the line order. But within a single row, boxes are not sorted left-to-right — they come out in roughly descending-side-length order, which (because longer detections tend to be the cleaner ones) reads as "sorted by probability."

Where it happens

packages/react-native-executorch/common/rnexecutorch/models/ocr/utils/DetectorUtils.cpp — groupTextBoxes:

The merge pass at line 544 sorts boxes descending by maxSideLength:

std::ranges::sort(boxes, [](const types::DetectorBBox &lhs,
                            const types::DetectorBBox &rhs) {
  return maxSideLength(lhs.bbox) > maxSideLength(rhs.bbox);
});

The final sort at line 605 then only compares the top-Y coordinate:

std::ranges::sort(mergedVec, [](const auto &obj1, const auto &obj2) {
  const float minY1 = minimumYFromBox(obj1.bbox);
  const float minY2 = minimumYFromBox(obj2.bbox);
  return minY1 < minY2;
});

For boxes with near-equal minY (i.e. same visual row), the order falls back to mergedVec insertion order, which is the side-length order from earlier. Result: longer boxes appear first within a row.

Reproduction

1. Open the computer-vision demo → OCR screen.
2. Pick a photo with multiple short tokens on the same line (e.g. a label with Price 12.99 USD).
3. Inspect the results list — boxes within the row come out re-ordered, not left-to-right.

Suggested fix

Row-bucket then sort by X. Two viable shapes:

• Soft compare: in the final sort comparator, treat |minY1 - minY2| < yThresh as same row and fall back to comparing minX. yThresh can be a fraction of the average detected box height.
• Explicit grouping: group boxes whose Y-extents overlap by ≥ N% into rows, sort rows by top-Y, then sort each row by min-X.

Notes

• Same code on main — unrelated to PR feat(constants)!: switch URLs to v0.9.0 layout + add MODEL_REGISTRY #1148 (model registry).
• Reported while testing the model-registry PR.