Modern object detectors can predict keypoints as per-object geometric features, but keypoints are inherently vulnerable to occlusion and truncation. Object edges can provide complementary information, often remaining observable even when partially occluded or when their defining endpoints leave the field of view. We therefore introduce object keyline detection, the task of predicting a fixed, semantically ordered set of infinite lines per object instance. Unlike the outputs of general line-segment detectors, each keyline is associated with an object instance and identified by a semantic index, providing a line-level analog of keypoints. To solve this task, we extend YOLO with keyline prediction parameterized in continuous Hesse-normal form [ρ,cosθ,sinθ], and introduce Object Keyline Similarity (OKLS), which serves as an evaluation metric, target assignment signal, and regression loss. We instantiate this formulation on runway perception in vision-based landing, a motivating example containing semantic linear features that remain partially observable throughout the approach. Experiments show that our representation achieves the highest detection accuracy among four evaluated line parameterizations, and keylines integrate directly into standard YOLO heads as additional output channels without degrading keypoint performance or adding meaningful latency.
Figures & tables
Figure 1: A partially visible runway during final approach. The two near corners are truncated by the image boundary, preventing keypoint-based pose estimation, while four of five runway keylines remain observable, providing line constraints for pose estimation.
Figure 2: The basic components of aircraft pose estimation for vision-based landing. An object detection algorithm detects a set of runway objects and their predetermined features, while a subsequent pose estimation algorithm estimates the aircraft’s pose. In this work we extend the detected feature set by keylines and evaluate them in the image domain.
Figure 3: Our proposed Object Keyline Similarity (OKLS)uses the orthogonal distances p1 and p2 from ground-truth support endpoints (e1,e2) to the predicted line l^ as its residual.
Figure 4: Head architecture variants for joint keypoint and keyline prediction: (a) Split: uses a dedicated keyline branch; (b) Fused: shares a single geometric branch across tasks.
Representation
Loss
APKL↑
ARKL↑
Continuous Hesse Normal
OKLS
55.7 (0.8)
61.7 (0.7)
Polar Hesse Normal
OKLS
53.0 (1.0)
59.5 (0.4)
Endpoints
OKLS
52.3 (1.2)
58.9 (1.4)
Endpoints
OKS
47.5 (0.8)
53.8 (1.0)
Normal Point
OKLS
52.6 (1.0)
59.2 (0.8)
Normal Point
OKS
46.5 (0.5)
52.1 (1.3)
Table 1: Ablation results comparing proposed and alternative line representations and loss functions. Values are mean (sample standard deviation) over three training runs with different seeds.
Figure 5: Keyline predictions on a two-runway scene (San Francisco International Airport). The zoomed view (right) shows all five predicted keylines per runway with accurate geometric alignment, illustrating the multi-instance prediction capabilities of our method.
Figure 6: Keyline predictions under nighttime conditions with truncated runway edges. The model recovers all visible keylines, although night scenes are underrepresented in training.
Arch.
APKL↑
ARKL↑
APKP↑
ARKP↑
Lat. [ms] ↓
Original
N/A
N/A
48.4 (0.5)
51.5 (0.4)
27.3 (0.3)
Fused
52.3 (0.9)
59.1 (0.9)
48.7 (0.8)
52.5 (0.2)
27.4 (0.2)
Split
52.0 (1.4)
59.3 (1.2)
49.8 (0.2)
52.8 (0.3)
29.2 (0.2)
Table 2: Joint keypoint and keyline prediction results (mean with sample standard deviation over three training runs). “Original” is the YOLO-NAS-Pose baseline trained only for runway-corner keypoints. “Fused” predicts keypoints and keylines through a shared geometric head, whereas “Split” uses a separate geometric branch for keyline regression ( Sec. 3.4 ). The joint models achieve comparable keypoint AP/AR to the keypoint-only baseline. While Fused and Split have comparable keyline performance, Fused incurs lower latency.
Model Configuration
APnPL↑
APnPL,4L↑
Ours (Keypoints + Keylines)
99.3%
99.0%
Lines-only
98.8%
97.1%
Original (Keypoints-only)
71.3%
N/A
Table 3: PnPL constraint availability ( APnPL ) across model configurations, measuring the detection of at least three features (which is a necessary condition for pose estimation). APnPL uses all predicted features, while APnPL,4L excludes the centerline. “Ours” refers to the proposed fused head model.
Figure 7: Common failure modes of unstructured line detectors. M-LSD (left) produces a duplicate centerline prediction. LINEA (middle) misses the far runway edge, but produces additional unwanted predictions for a taxiway with strong linear features on the right side. Neither model predicts a fixed semantic keyline set; our method (right) correctly produces predictions for the five defined runway lines.
Model
APKL↑
ARKL↑
Latency [ms] ↓
M-LSD Tiny
10.7
26.3
5.6
M-LSD Large
11.0
26.7
34.6
LINEA Small
39.6
58.7
43.4
LINEA Medium
42.4
58.8
58.9
Ours Nano
55.7
61.7
27.4
Ours Medium
59.1
65.7
81.1
Table 4: Comparison of our proposed method (keylines-only) with general line detection models evaluated on semantic keyline benchmark metrics. Best values are bold. Latencies as in Sec. 4.4 .