When runway corners get cropped, YOLO learns whole edges instead: five semantic lines keep 99.3% of approach images constrained
Synopsis
In vision-based landing, runway corner keypoints are often cropped by the field of view during close approach, leaving point-based pose solvers without geometric constraints. The work extends object detection from keypoints to keylines: each object instance predicts a fixed, semantically ordered set of infinite lines, implemented on YOLO with a continuous Hesse normal parameterization and Object Keyline Similarity (OKLS) serving as evaluation metric, target assignment signal and regression loss. On runway data, the continuous Hesse form with OKLS gives the highest keyline detection accuracy among four parameterizations, and a fused head adds keylines without degrading keypoint accuracy or meaningfully increasing latency.
Interpretation
Object detectors can now predict, alongside keypoints, a fixed, semantically ordered set of infinite lines per object instance, called keylines. General line-segment detectors output unordered, image-level finite segments that are neither tied to an object instance nor given a stable semantic identity; object detectors give boxes and keypoints but no line features. On runway approach images, keyline detection with the continuous Hesse normal form and OKLS supervision reaches 55.7 AP and 61.7 AR, above the polar Hesse form, endpoint representation and normal-point representation under the same YOLO-NAS architecture.
Each keyline is represented in continuous Hesse normal form, a unit normal vector plus a signed perpendicular distance, and decoded relative to the output grid cell center. Regressing a scalar angle directly is discontinuous where the angle wraps around, so the model predicts a unit normal instead, at the cost that opposite normals describe the same undirected line. OKLS replaces the point-to-point residual of keypoint similarity with the perpendicular distance from the predicted line to the annotated support endpoints, and is used for evaluation, target assignment and regression loss, yielding keyline AP and AR under the COCO keypoint protocol.
Keylines plug directly into standard YOLO keypoint heads as additional output channels without substantial architectural change. Both fused-head and split-head integrations predict keypoints and keylines; the fused head shares the existing pose branch, while the split head duplicates the geometric regression branch. The fused head gives 48.7 AP and 52.5 AR for keypoints, comparable within run-to-run variation to the keypoints-only model's 48.4 AP and 51.5 AR, at 27.4 ms versus 27.3 ms latency; the split head adds 1.9 ms.
In the runway setting, keylines raise the share of images that supply usable feature constraints to downstream pose solving. The keypoints-only model has at least three visible keypoints in only 71.3% of test images, since once the two nearest corners leave the frame at most two corner keypoints remain. Keypoints plus keylines gives at least three valid features in 99.0% of test images and the lines-only configuration in 97.1%; excluding the centerline, the lines-only and joint models retain their respective availability, so the gain is not explained by the five-versus-four feature count alone.
Perspective
All experiments use a runway dataset of real approach images: 6730 training images from 230 approach trajectories at 95 airports, with validation and test sets each covering 10 airports disjoint from training. The method assumes instances of a category have a fixed, semantically ordered set of approximately linear features, which runways satisfy because their edges and centerline have stable semantic identities. The authors propose next integrating predicted keylines into PnPL pose estimation to evaluate their effect on aircraft pose accuracy during final approach, and validating the formulation on categories beyond runways.
Keyline regression is evaluated in the image domain only, not its effect on downstream pose estimation, so the results do not establish that predicted keylines improve aircraft pose accuracy. No geometric constraints such as a shared vanishing point are enforced on keylines, to avoid structural error coupling across features. Joint training lowers keyline accuracy by 3.4 to 3.7 points relative to the keyline-only model, which the authors attribute possibly to the constant multi-task loss weight. Constraint availability counts valid features in the image domain; reaching three correspondences is a necessary generic minimum but does not guarantee successful pose estimation, and geometric degeneracy is not tested.