OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping

1 Introduction

• Separate lane and traffic-signal datasets provide limited scene understanding, leaving ambiguity about lane guidance at intersections and which lanes a signal controls.The paper also notes that 2D lanelines can be insufficient for real-world requirements and that absent markings can leave vehicles without trajectory guidance.Source: §1 · pp. 1–2
• OpenLane-V2 unifies 3D lane centerlines, traffic elements, and topology relationships, with the OpenLane-V2 Score combining perception (DET) and reasoning (TOP).A lane–traffic-element correspondence is valid only when the traffic element controls that lane.Source: §1 · p. 2
• OpenLane-V2 is presented as the first autonomous-driving dataset focused on topology reasoning, with dedicated tasks and metrics for benchmarking scene understanding.The introduction says the benchmark builds on existing benchmarks and includes images collected from cities worldwide.Source: §1 · p. 3
• The dataset includes 2.1M instance-level annotations and 1.9M positive topology relationships, with annotations carefully validated.A development kit, framework plug-ins, test server, and leaderboard are also described as maintained resources.Source: §1 · p. 3

2.1 3D Lane Detection

• OpenLane-V2 extends 3D lane annotation beyond visible front-view lanelines to conceptual centerlines around the vehicle, providing trajectory guidance in all directions.Earlier lane datasets largely label 2D lanelines; OpenLane was the first large-scale, real-world 3D laneline dataset.Source: §2.1 · p. 3

2.2 Traffic Element Recognition

• Earlier traffic-element datasets mainly emphasize object location, whereas OpenLane-V2 adds topology relationships between traffic elements and lanes.The table compares datasets by images, classes, annotations, tracking, and whether lane–traffic-element correspondences are annotated.Source: §2.2 · p. 4, Table 1 · p. 4

2.3 Scene Understanding

• Prior scene-understanding datasets and relationship tasks cover limited 2D object sets or movable-agent behavior, while this work emphasizes static-scene understanding that provides trajectory information.The related work contrasts this focus with datasets on human-object interactions and on the intentions or future behavior of agents.Source: §2.3 · p. 4

3 OpenLane-V2

• The dataset uses 2,000 scene segments from Argoverse 2 and nuScenes, with high-quality annotations produced through experienced annotators and multiple validation stages.Scenes cover varied cities and conditions, including noon and night, sunny and rainy days, downtown and suburbs.Source: §3 · p. 4

3.1 Raw Data Acquisition

• Two subsets reflect the source datasets’ different sensor setups, with a unified right-handed coordinate system and camera and ego-pose parameters provided.subset_A contains scenes from six cities and subset_B from Boston and Singapore; their night and rain proportions differ.Source: §3.1 · p. 5

3.2 Centerlines and Their Connectivity

• Lane centerlines are regressed from HD-map boundaries, merged across single-predecessor or single-successor lanes, and split at intersections, forks, and merges.Each centerline is an ordered list of 3D points; subset_B’s z values are set to 0 because its HD maps lack height information. Lane direction runs from start to end, and connectivity links one lane’s endpoint to another’s start.Source: §3.2 · p. 5

3.3 Traffic Elements and Their Correspondence to Centerlines

• Traffic elements receive front-view 2D boxes and semantic attributes, while composite attributes are split into separate annotations sharing the same box.There are 13 attributes; elements with unobservable meaning are labeled unknown. U-turn annotations are much rarer than common red or green traffic lights.Source: §3.3 · p. 6
• Traffic-element correspondence is assigned by lane-driving rules: nondirectional elements associate with lanes permitting straight travel, and directional elements with lanes moving in the indicated direction.Traffic elements are associated only with centerlines outside intersections.Source: §3.3 · p. 6

4 Task Definition & Evaluation Metric

• The OpenLane-V2 Score averages metrics from 3D lane detection, traffic-element recognition, and topology recognition to describe primary-task performance.The primary task includes recognizing lanes and their dynamic drivable states, traffic-element attributes, and relationships among perceived entities.Source: §4 · p. 6

4.1 3D Lane Detection

• 3D lane detection perceives directed 3D lane centerlines from multi-view images covering a panoramic field of view.Centerline similarity uses discrete Fréchet distance, and DET_l is averaged over match thresholds {1.0, 2.0, 3.0}; thresholds are relaxed for distant lanes based on distance from the ego car.Source: §4.1 · pp. 6–7

4.2 Traffic Element Recognition

• Traffic-element recognition detects front-view locations and attributes together, with tiny element size identified as a challenge in large outdoor scenes.The evaluation uses IoU distance with a match threshold of 0.75 and averages DET_t over attributes.Source: §4.2 · p. 7

4.3 Topology Reasoning

• Topology reasoning evaluates directed lane connectivity and lane–traffic-element correspondence as separate graphs.TOP is an mAP metric on matched graphs, using entity-specific Fréchet and IoU similarities to project predicted vertices; TOP_ll and TOP_lt score lane connectivity and lane-to-traffic topology, respectively.Source: §4.3 · p. 7

5 Experiments

• The experiments adapt and evaluate multiple state-of-the-art methods, using visualizations and analysis to examine design choices.The qualitative results suggest that adding traffic-rule priors could be a potential way to improve reasoning about which traffic elements are associated with a lane.Source: §5 · p. 8, §5.2 · p. 8

5.1 Baselines

• The baselines include methods for centerline connectivity, point-sequence or point-set map prediction, and joint scene-structure understanding.STSU derives connectivity from detected centerlines; VectorMapNet predicts point sequences autoregressively; MapTR uses equivalent point-set permutations; TopoNet uses instance-level queries for centerlines and traffic elements.Source: §5.1 · p. 8

5.2 Results

• Existing methods struggle with centerline detection and connectivity, while tiny traffic elements also make recognition difficult.TOP requires matching predicted and ground-truth entities, so weak centerline perception also limits connectivity performance.Source: §5.2 · p. 8
• At a complex crossroad, all models approximate the intersection, but centerline positions, shapes, and topology remain imperfect.STSU produces many false-positive centerlines and fails to attach traffic information; MapTR's independent lane points do not form smooth driving trajectories, whereas VectorMapNet's autoregressive design produces more properly shaped centerlines. All three of VectorMapNet, MapTR, and TopoNet provide only partially correct centerline–traffic-element associations.Source: §5.2 · pp. 8–9

6 Conclusion

• OpenLane-V2 represents lane networks with centerlines and connectivity, and traffic information with semantic labels and lane associations.The paper describes tasks and metrics and evaluates various methods on the dataset.Source: §6 · p. 9
• The dataset builds on existing datasets and remains at the same scale as previous datasets; the authors expect more driving scenes to increase diversity.Planning is not included because it also requires movable objects such as cars and pedestrians for collision avoidance; integrating static and dynamic entities is left for future work.Source: §6 · p. 9
• Models trained or evaluated on the dataset should not be used directly for deployment or real-world applications, and the dataset gives no safety-critical guarantees.Source: §6 · p. 9

A Overview

• The supplementary material includes author statements, licensing, benchmark implementation details, and Datasheet-format documentation.The documentation covers dataset distribution, maintenance, composition, and collection.Source: A Overview · p. 13

B Author Statement

• The authors accept responsibility for licensing, distribution, and maintenance of the dataset.Source: B Author Statement · p. 13

C Licensing

• The dataset uses CC BY-NC-SA 4.0, while the repository code uses the Apache License 2.0.Source: C Licensing · p. 13

D.1 Motivation

• The dataset aims to depict complex traffic scenes with instance and topology annotations, including directed centerlines, lane connectivity, semantic traffic elements, and lane-control associations.The datasheet says centerlines offer trajectory guidance and that traffic-element associations indicate lanes controlled under traffic rules.Source: §D.1 · p. 13
• The dataset was created by researchers from OpenDriveLab and Huawei, with funding from Shanghai AI Laboratory and Huawei.Source: §D.1 · p. 13

D.2 Distribution

• The dataset is publicly accessible online and is distributed through the OpenDriveLab OpenLane-V2 GitHub repository.Source: §D.2 · p. 13

D.3 Maintenance

• The authors support, host, and maintain the dataset; they plan to post updates, continue hosting older versions, and provide a mechanism for contributions.Source: §D.3 · p. 14

D.4 Composition

• Each frame identifies lane centerlines, traffic elements, centerline connectivity, and which traffic signals control particular lanes.A vehicle must follow a signal if and only if a topology relationship connects it to its lane.Source: §D.4 · p. 14
• The dataset contains 2,000 scenes and 72K frames, with about 1.8 million lane-centerline and 0.25 million traffic-element annotations.Frames are annotated at 2Hz, and each scene contains a varying number of consecutive frames.Source: §D.4 · p. 14
• Connectivity between lanes and relationships between lanes and traffic elements are supplied as per-frame adjacency matrices.Frames within a segment are consecutively sampled at 2Hz.Source: §D.4 · p. 14
• The dataset is partitioned into training, validation, and testing splits and redistributes reconstructed raw data from Argoverse 2 and nuScenes.The authors describe it as self-contained despite its source datasets.Source: §D.4 · p. 14

D.5 Collection Process

• Experienced annotators and multiple validation stages support annotation quality; contractors make and verify traffic-element and topology annotations.Source: §D.5 · p. 14

D.6 Use

• The dataset supports 3D lane detection, traffic-element recognition, and topology reasoning.Source: §D.6 · p. 15

• For a vertex, the example ranks predicted neighbors by confidence against its ground-truth connected neighbors to calculate mAP.The example uses vertex 1, whose ground-truth neighbors are {2, 3, 5, 8}, and predicted neighbors are {4, 3, 5, 6, 7}.Source: E Evaluation Metric · p. 15
• The TOP calculation combines a correctness indicator for each predicted connection with the precision of predicted neighbors, then averages vertex mAPs.Source: E Evaluation Metric · p. 15
• The reported value is 7/192, approximately 0.036.Source: E Evaluation Metric · p. 15

F Implementation Details

• STSU, VectorMapNet, and MapTR are adapted to the benchmark alongside TopoNet, a method designed for scene-structure understanding.Source: F Implementation Details · p. 15
• All models use consistent input settings and are trained on both subsets; a typical 24-epoch run takes about 16 hours on eight NVIDIA A100 GPUs.Subset A images are 2048 × 1550 except the front view; subset B images are 1600 × 900.Source: F Implementation Details · p. 15
• TopoNet uses a ResNet-50 backbone, C3–C5 features, and a 200 × 100 BEV grid, with specified loss weights for traffic-element and lane-centerline tasks.Source: F Implementation Details · p. 15
• STSU predicts centerlines with a DETR head and connectivity with decoder queries; its adaptation concatenates multi-view BEV embeddings and adds TopoNet-style traffic-element and lane-topology heads.Bezier control points are interpolated into 11 equidistant points.Source: F Implementation Details · p. 15
• VectorMapNet uses decoder keypoints and autoregressive position generation; its benchmark adaptation predicts 2D BEV centerlines and omits height during evaluation.Its perception range is expanded to ±50m × ±25m, and centerlines are interpolated to match the benchmark setting.Source: F Implementation Details · p. 16
• MapTR uses averaged point queries as centerline instance queries and predicts 2D BEV centerlines, disregarding z during inference.The adaptation increases perception range to ±50m × ±25m and uses 11 point queries per lane.Source: F Implementation Details · p. 16

G Illustrations

• Figure 1 contrasts a red-light-controlled left-turn lane with a straight lane that can continue, while Figure 2 contrasts centerline attributes with separate laneline and road-element detection plus hand-crafted rules.Source: Figure 1 · p. 16, Figure 2 · p. 17
• Figures 3–5 depict centerline segmentation at lane splits and merges, a traffic element associated with three centerlines, and numerous centerlines in intersections.Source: Figure 3 · p. 17, Figure 4 · p. 18, Figure 5 · p. 18