GLane3D: Detecting Lanes with Graph of 3D Keypoints

Figure: Overview of the GLane3D model architecture.

Abstract

Accurate and efficient lane detection in 3D space is essential for autonomous driving systems, where robust generalization is the foremost requirement for 3D lane detection algorithms. Considering the extensive variation in lane structures worldwide, achieving high generalization capacity is particularly challenging, as algorithms must accurately identify a wide variety of lane patterns worldwide. Traditional top-down approaches rely heavily on learning lane characteristics from training datasets, often struggling with lanes exhibiting previously unseen attributes. To address this generalization limitation, we propose a method that detects keypoints of lanes and subsequently predicts sequential connections between them to construct complete 3D lanes. Each key point is essential for maintaining lane continuity, and we predict multiple proposals per keypoint by allowing adjacent grids to predict the same keypoint using an offset mechanism. PointNMS is employed to eliminate overlapping proposal keypoints, reducing redundancy in the estimated BEV graph and minimizing computational overhead from connection estimations. Our model surpasses previous state-of-the-art methods on both the Apollo and OpenLane datasets, demonstrating superior F1 scores and a strong generalization capacity when models trained on OpenLane are evaluated on the Apollo dataset, compared to prior approaches.

Method Overview

(a)

(b)

(c)

(d)

a) Multiple Proposals: Multiple keypoint proposals are generated for each lane by selecting the top N anchor points with the highest scores from the predicted segmentation map 𝑀_seg. These proposals are sampled within a lateral distance dₓ from the target lanes, increasing robustness to ensure that no keypoints are missed due to sparse or noisy predictions.
b) Keypoint Refinement: Each proposed keypoint is refined toward its corresponding location on the lane by applying a predicted lateral offset. This adjustment aligns proposals more precisely to the actual lane geometry, improving spatial accuracy and continuity along the lane.
c) Redundancy Suppression (PointNMS): To reduce redundancy and computational load, PointNMS is applied to suppress overlapping proposals. Only the top S keypoints with the highest confidence scores are retained, ensuring that the remaining keypoints are both distinct and reliable.
d) Connection Estimation: Directed connections between the retained keypoints are estimated to construct a lane graph. A relationship head predicts an adjacency matrix 𝑨, encoding which keypoints are connected in a sequential manner to represent continuous lane segments.
Lane Extraction: Final lanes are extracted by applying Dijkstra’s shortest path algorithm over the graph of keypoints. This graph-based decoding traces the most plausible paths between start and end keypoints, enabling robust reconstruction of complete lane topologies.

Quantitative Results @ 1.5m (OpenLane Dataset)

Method	Backbone	F1-Score ↑	X-error near (m) ↓	X-error far (m) ↓	Z-error near (m) ↓	Z-error far (m) ↓
PersFormer	EffNet-B7	50.5	0.485	0.553	0.364	0.431
BEV-LaneDet	ResNet-34	58.4	0.309	0.659	0.244	0.631
Anchor3DLane	EffNet-B3	53.7	0.276	0.311	0.107	0.138
LATR	ResNet-50	61.9	0.219	0.259	0.075	0.104
LaneCPP	EffNet-B7	60.3	0.264	0.310	0.077	0.117
PVALane	ResNet-50	62.7	0.232	0.259	0.092	0.118
PVALane	Swin-B	63.4	0.226	0.257	0.093	0.119
GLane3D-Lite (Ours)	ResNet-18	61.5	0.221	0.252	0.073	0.101
GLane3D-Base (Ours)	ResNet-50	63.9	0.193	0.234	0.065	0.090
GLane3D-Large (Ours)	Swin-B	66.0	0.170	0.203	0.063	0.087

Quantitative Results @ 0.5m (OpenLane Dataset)

Method	Backbone	F1-Score ↑	X-error near (m) ↓	X-error far (m) ↓	Z-error near (m) ↓	Z-error far (m) ↓
PersFormer	EffNet-B7	36.5	0.343	0.263	0.161	0.115
Anchor3DLane	EffNet-B3	34.9	0.344	0.264	0.181	0.134
PersFormer	ResNet-50	43.2	0.229	0.245	0.078	0.106
LATR	ResNet-50	54.0	0.171	0.201	0.072	0.099
DV-3DLane(Camera)	ResNet-34	52.9	0.173	0.212	0.069	0.098
GLane3D-Lite (Ours)	ResNet-18	53.8	0.182	0.206	0.070	0.095
GLane3D-Base (Ours)	ResNet-50	57.9	0.157	0.179	0.067	0.087
GLane3D-Large (Ours)	Swin-B	61.1	0.142	0.167	0.061	0.084

Cross-Dataset Evaluation on Apollo Balanced Scenes

Table: Cross-Dataset Evaluation on Apollo Balanced Scenes
Dₜ	Methods	F1	X-error (m) ↓		Z-error (m) ↓
Dₜ	Methods	F1	Near	Far	Near	Far
1.5m	PersFormer	53.2	0.407	0.813	0.122	0.453
	LATR	34.3	0.327	0.737	0.142	0.500
	GLane3D	58.9	0.289	0.701	0.086	0.479
0.5m	PersFormer	17.4	0.246	0.381	0.098	0.214
	LATR	19.0	0.201	0.313	0.116	0.220
	GLane3D	42.6	0.162	0.296	0.063	0.198

Apollo Cross-Dataset Comparison — Figure: Visual comparison on Apollo dataset.

Apollo Dataset Results

Table: Quantitative Results on the Apollo 3D Synthetic Dataset
Subset	Methods	Backbone	F1-Score (%) ↑	AP (%) ↑	X-error near (m) ↓	X-error far (m) ↓	Z-error near (m) ↓	Z-error far (m) ↓
Balanced Scene	PersFormer	EffNet-B7	92.9	-	0.054	0.356	0.010	0.234
	BEVLaneDet	ResNet-34	96.9	-	0.016	0.242	0.020	0.216
	LaneCPP	EffNet-B7	97.4	99.5	0.030	0.277	0.011	0.206
	LATR	ResNet-50	96.8	97.9	0.022	0.253	0.007	0.202
	DV-3DLane	ResNet-50	96.4	97.6	0.046	0.299	0.016	0.213
	GLane3D (Ours)	ResNet-50	98.1	98.8	0.021	0.250	0.007	0.213
Rare Scene	PersFormer	EffNet-B7	87.5	-	0.107	0.782	0.024	0.602
	BEVLaneDet	ResNet-34	97.6	-	0.031	0.594	0.040	0.556
	LaneCPP	EffNet-B7	96.2	98.6	0.073	0.651	0.023	0.543
	LATR	ResNet-50	96.1	97.3	0.050	0.600	0.015	0.532
	DV-3DLane	ResNet-50	95.5	97.2	0.071	0.664	0.025	0.568
	GLane3D (Ours)	ResNet-50	98.4	99.1	0.044	0.621	0.023	0.566
Visual Variations	PersFormer	EffNet-B7	89.6	-	0.074	0.430	0.015	0.266
	BEVLaneDet	ResNet-34	95.0	-	0.027	0.320	0.031	0.256
	LaneCPP	EffNet-B7	90.4	93.7	0.054	0.327	0.020	0.222
	LATR	ResNet-50	95.1	96.6	0.045	0.315	0.016	0.228
	DV-3DLane	ResNet-50	91.3	93.4	0.095	0.417	0.040	0.320
	GLane3D (Ours)	ResNet-50	92.7	94.8	0.046	0.364	0.020	0.317

OpenLane Qualitative Results

More Cross-Dataset Qualitative Results

BibTeX

@inproceedings{ozturk2025glane3d,
          title={GLane3D: Detecting Lanes with Graph of 3D Keypoints},
          author={{\"O}zt{\"u}rk, Halil {\.I}brahim and Kalfao{\u{g}}lu, Muhammet Esat and Kilinc, Ozsel},
          booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
          pages={27508--27518},
          year={2025}
        }