High-definition maps provide critical structural information for downstream autonomous driving tasks. Recent methods increasingly rely on external priors to improve prediction accuracy and robustness. Yet, existing prior representations expose a trade-off: dense rasters are easy to integrate but ignore certain map geometric properties, while sparse vectors preserve geometric structure but require dedicated encoders and are difficult to plug into existing architectures. Moreover, external priors can be inaccurate, costly to acquire, or difficult to maintain, motivating a structural representation that can provide geometric guidance while remaining learnable and broadly applicable. We address this limitation with the Lane Field, a dense BEV-aligned geometric representation that encodes local lane structure at each spatial location and can be either predicted from onboard observations or constructed from external map priors. Building on this representation, we introduce a plug-and-play framework that enhances online HD mapping through two complementary mechanisms: BEV feature modulation, which injects lane-structured cues into latent BEV features, and deformable attention guidance, which steers query sampling toward the underlying lane geometry. We primarily evaluate our method on OpenLane-V2, showing that conditioning BEV features and guiding deformable attention with local lane geometry enables our framework to consistently improve existing methods and achieve state-of-the-art performance on lane segment perception. Project page: https://andberto.github.io/LaneField/
Figures & tables
Figure 1: Comparison between classical map representations commonly used as additional supervision or external priors and our proposed representation, the Lane Field. Our method elegantly combines the geometric meaning of vectorized representations with a dense BEV-aligned spatial structure.
Figure 2: Overview of the proposed framework. Starting from FBEV , a prediction head estimates the Lane Field (a) , supervised by ground-truth fields geometrically constructed from vectorized map annotations. The predicted LF is then used to modulate the BEV features through a SPADE-like residual block (b) and to guide deformable attention sampling (c) toward geometrically meaningful roadway regions.
Figure 3: (a) SPADE-like modulation block. Given the predicted Lane Field, we project it into a shared latent representation, which is then processed by two separate convolutional branches to predict the modulation parameters γ and β . (b) Residual BEV modulation block. The modulation operation is embedded into a residual block, which can be stacked to progressively refine the BEV features.
Figure 4: Effect of the Lane Field guidance mechanism on deformable attention sampling locations around a reference point. When all three gates are active, the sampling locations are steered along the local lane direction, yielding a lane aligned sampling pattern. The tangential stretch is defined in Equation 12 .
Table 1: Comparison on OpenLane-V2 subset A. (a) Results on lane segment perception. (b) Results on centerline perception, where OLS l is computed using only centerline detection and lane-lane topology. All models adopt ResNet-50 as backbone and are trained for 24 epochs. † : Denotes the baseline trained from the official public implementation. ‡ : Denotes methods that exploit temporal information. ∗ : TopoFormer uses a pre-trained lane detector. The best and second best results are highlighted in bold and underlined , respectively, while “–” indicates no relevant results. All the deltas are reported with respect to the corresponding baseline.
Component
mAP ↑
AP ls ↑
AP ped ↑
TOP lsls ↑
Baseline
32.1
32.2
31.9
25.5
+ Guidance
33.2 (+1.1)
33.4 (+1.2)
32.9 (+1.0)
26.8 (+1.3)
+ Modulation
34.9 (+2.8)
35.7 (+3.5)
34.0 (+2.1)
27.9 (+2.4)
+ Both
36.1 (+4.0)
35.8 (+3.6)
36.4 (+4.5)
28.4 (+2.9)
Table 2: Component-wise ablation on OpenLane-V2 with LaneSegNet, evaluating BEV feature modulation, deformable attention guidance, and their combination.
Method
mAP ↑
AP ls ↑
AP ped ↑
TOP lsls ↑
LaneSegNet
32.1
32.2
31.9
25.5
+ Ours (SD)
36.8 (+4.7)
36.6 (+4.4)
37.0 (+5.1)
29.7 (+4.2)
Δrel
+14.6%
+13.7%
+16.0%
+16.5%
Table 3: Comparison between the LaneSegNet baseline and the variant using SD maps derived lane fields on OpenLane-V2.
Figure 5: Qualitative results on OpenLane-V2. We show the predicted map reprojected onto the perspective-view images and the corresponding BEV comparison among the ground truth, LaneSegNet baseline, and our Lane Field enhanced variant.
Dataset
Method
mAP ↑
AP div ↑
AP ped ↑
AP bou ↑
AP cen ↑
AV2
MapTRv2
62.0
66.3
54.8
64.6
62.0
+ Ours
64.2
69.5
56.4
67.1
63.6
GeMap
70.3
74.6
66.7
69.5
–
+ Ours
71.2
75.3
67.4
70.8
–
NUS
MapTRv2
55.8
55.9
52.5
60.1
54.9
+ Ours
57.7
59.2
52.9
62.8
55.8
Table 4: Analysis on Argoverse 2 and NuScenes using MapTRv2 and GeMap as baselines for the standard map element detection.
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 1: Effect of the spread factor f on the generated ground truth Lane Field. From left to right, fields are generated with f=0.25 , f=0.50 , and f=0.75 .
f
σi
Weight at lane edge
0.25
wˉi/4
13%
0.50
wˉi/2
61%
0.75
0.75wˉi
80%
1.00
wˉi
88%
Appendix
Table 1: Effect of the scale factor f on the Gaussian weight at the lane edge, i.e., at distance wˉi/2 from the centerline.
Figure 2: Visualization of the spatial gradients of the ground truth tensor channels along the horizontal ( x ) and vertical ( y ) directions. The first row shows Δx for Mxx , Myy , and Mxy , while the second row shows Δy for the same channels.
Parameter
Value
Description
wR
5.0
LR loss weight
wG
2.0
LG loss weight
wL
1.0
LL loss weight
β
0.5 from [ 9 , 3 ]
β -NLL β param.
wmin
0.1
Background weight
τ
0.2
Prox. gate threshold
Appendix
Table 2: Hyperparameters used for lane field supervision. Values are selected empirically based on validation performance.
Gate
Parameters
Anisotropy gate
αC=10.0,τC=0.3
Uncertainty gate
αU=10.0,τU=0.3
Learnable per-head gate
αH=2.0
Max Prior Stretch
Δ∥=4.0
Appendix
Table 3: Parameters of the gates used in the deformable attention guidance mechanism.
Method
Ealign↓
Anisotropy
Area
Lane Attention [ 4 ]
35.3°
13%
33.2 m2
Guided Attention ( Ours )
11.4°
62%
78.2 m2
Appendix
Table 4: Quantitative comparison between the original Lane Attention of LaneSegNet [ 4 ] and our conditioned variant.
Baseline
Model
+SPADE
+Attn.
+All
LaneSegNet [ 4 ]
14.6 / 1.3
13.9 / 1.5
13.9 / 1.5
13.9 / 1.5
MapTRv2 [ 5 ]
33.2 / 1.2
29.9 / 1.4
29.9 / 1.4
29.9 / 1.4
TopoPoint [ 1 ]
13.8 / 1.4
13.0 / 1.6
13.0 / 1.6
13.0 / 1.6
Appendix
Table 5: Breakdown of the computational overhead introduced by the proposed Lane Field modules across different baselines. Results are reported as FPS / TFLOPs, respectively.
Baseline
Model
+Head
+SPADE
+Attn.
+All
LaneSegNet [ 4 ]
45.4M
4.8M
2.4M
≈ 0
52.6M
MapTRv2 [ 5 ]
40.6M
4.8M
2.4M
≈ 0
47.8M
TopoPoint [ 1 ]
53.6M
4.8M
2.4M
≈ 0
60.8M
Appendix
Table 6: Breakdown of the parameter overhead introduced by the proposed Lane Field modules across different baselines.
Gradients conf.
mAP ↑
AP ls ↑
AP ped ↑
TOP lsls ↑
Baseline
32.1
32.2
31.9
25.5
Stop-grad. In & Out
34.1
34.2
34.0
26.7
Stop-grad. In
34.8
34.4
35.2
27.0
Stop-grad. Out
33.3
32.3
34.3
27.2
Fully attached
36.1
35.8
36.4
28.4
Appendix
Table 7: Gradient-flow configurations of our framework. Stop-grad. In detaches the BEV features before the field head, while Stop-grad. Out detaches the LF at the head output.
Figure 3: Predicted variance as an uncertainty cue for lane field guidance. The driveway is not annotated in the ground truth (left). Although the model predicts high proximity in that region (center), it also predicts high variance (right), indicating uncertainty and allowing downstream modules to reduce the use of the lane field there.
Figure 4: The parking lot entrance is not annotated in the ground truth (left). Although the model predicts high proximity in that region (center), it also predicts high variance (right), indicating uncertainty and allowing downstream modules to reduce the use of the lane field there.
Figure 5: Example of an ambiguous free-space region outside the drivable lanes. The model assigns high variance to regions where it is uncertain whether the observed area should be treated as drivable area.
Figure 6: Visualization of the sampling locations associated with a reference point lying on the road at the last decoder layer. Without guidance (left), the sampling locations follow a mostly local and isotropic pattern, and the first principal component, i.e . the direction of maximum variance is misaligned with the local lanes direction. With lane field guidance (center), the sampling pattern is stretched along the lane structure, and its first principal component becomes well aligned with the local lane direction. The right panel compares the original sampling locations with the guided ones, showing how the offsets are stretched along the local lane direction.
Figure 7: Visualization of the sampling locations associated with a reference point lying at an intersection in the last decoder layer. Since the direction confidence C(p) is close to zero in this ambiguous region, the lane field guidance is suppressed and the sampling pattern remains close to the original learned offsets.
Figure 8: Visualization of the sampling locations associated with a reference point lying in a high variance region, corresponding to the area highlighted in Figure 4 , at the last decoder layer. Since the variance is high in this region, the uncertainty gate suppresses the lane field guidance and the deformable attention sampling remains close to the original learned offsets. The right panel highlights in red the areas considered reliable by the uncertainty gate.
Figure 9: Alignment error (in degrees) between the principal axis of the sampling locations, i.e., the direction of maximum spatial extent, and the ground-truth lane direction. Our conditioning method (red) exhibits substantially better alignment than the baseline sampling locations (orange).
Figure 10: Anisotropy comparison between our conditioned sampling locations (red) and the baseline sampling locations (orange). Our approach exhibits a clearer anisotropic pattern aligned with the lane direction, whereas the baseline sampling locations remain largely isotropic.
Figure 11: Area comparison (in m2 ) between our conditioned sampling locations (red) and the baseline sampling locations (orange). The coverage is computed as the average area covered per reference point. Our method achieves consistently larger spatial coverage than the baseline.
Figure 12: Comparison of BEV features before and after the proposed modulation block. The modulated features show a clearer road-structured organization, suggesting that the proposed lane field conditioning effectively enriches the BEV representation with geometric cues.
Figure 13: Qualitative comparison on OpenLane-V2. The LaneSegNet baseline fails to recover a pedestrian crossing and produces inaccurate local predictions, whereas our method correctly detects the crossing and yields a more precise and structurally consistent map.
Figure 14: Another comparison on OpenLane-V2 between the LaneSegNet baseline and our method. In this example, the baseline fails to recover a distant intersection visible in the HD map, whereas our method preserves its overall structure more accurately. We also report the predicted and ground truth lane field channels in BEV to illustrate the geometric cues used by our framework.
Figure 15: Failure case on OpenLane-V2. Due to the rarity of roundabouts in the dataset, both the LaneSegNet baseline and our method struggle to produce an accurate HD map prediction in this scenario. Nevertheless, our method shows a slightly better awareness of the overall lane structure.
Figure 16: Another failure case on OpenLane-V2. Although our method improves the overall lane structure prediction, the lane field does not explicitly encode pedestrian crossing regions, and the model may therefore still miss some crossings in challenging scenes.
Accurate, up-to-date representations of road structures are critical for the safe operation of autonomous vehicles. Existing systems rely either on costly, maintenance-heavy high-definition (HD) maps which compromise safety when outdated, or purely sensor-based online mapping which struggles with long-range reliability and occlusion. Systems incorporating map prior information into online mapping seek to overcome drawbacks of both approaches by combining them in some way. We propose 'Driveline To HD Map' (D2HDMap), an online mapping system that injects a lightweight, non-visible driveline prior to guide the estimation of visible road structures such as lane dividers, road boundaries and crosswalks. This prior incurs less effort to create and update compared to full HD map priors used in other approaches. We also show that training with such a prior can improve generalization at inference time when no prior is available. Ablation studies conducted on the nuScenes and Argoverse 2 dataset demonstrate that models trained using a driveline prior largely retain performance even when priors are not available. On a geographically disjoint split, D2HDMap achieves 44.8 mAP, surpassing recent state-of-the-art. Additionally, noise-aware training substantially increases robustness to realistic localization error.
Seojun Shon, Chikao Tsuchiya, Dhaval Bhanderi +3
Nissan North America · Nissan Advanced Technology Center - Silicon Valley, Nissan North America, Santa Clara, CA 95014, USA
Autonomous driving systems benefit from high-definition (HD) maps that provide critical information about road infrastructure. The online construction of HD maps offers a scalable approach to generate local vectorized maps from onboard sensor observations. Existing methods commonly adopt bird's-eye-view (BEV) features as the intermediate scene representation, encoding the surrounding space with fixed-resolution dense grids. However, map elements are spatially sparse yet require fine-grained geometric localization, making uniformly allocated BEV representations redundant and less effective for vectorized map prediction. In this work, we propose GaussianMap, an online HD map construction framework that learns an adaptive Gaussian representation of the surrounding scene. This representation consists of a set of Gaussian primitives on the BEV plane, each encoding a flexible local region with geometric properties and a feature vector, allowing the model to allocate representational capacity to map-relevant regions. To generate such a representation from sensor observations, we introduce a feed-forward Gaussian encoder that progressively refines these primitives through Gaussian interaction modeling and multi-sensor feature aggregation. The refined Gaussian representation is then splatted into a BEV feature map and decoded into vectorized map predictions. Extensive experiments on nuScenes and Argoverse 2 datasets demonstrate that GaussianMap achieves state-of-the-art performance in both camera-only and camera-LiDAR fusion settings. Our code will be made publicly available.
Hongyu Lyu, Julie Stephany Berrio Perez, Mao Shan +1
The University of Sydney, Australian Centre for Robotics, Sydney, Australia. · The University of Queensland, Australia.
Accurate High-Definition (HD) map construction is critical for autonomous driving, yet existing methods face a fundamental trade-off: vectorization-based approaches preserve topology but struggle with geometric fidelity, while rasterization-based approaches enable precise geometric supervision but produce unstructured outputs. To bridge this gap, we propose GSMap, a novel framework that unifies both paradigms via a learnable 2D Gaussian representation. Each map element is modeled as an ordered sequence of 2D Gaussians, whose centers correspond to the vertices of the vectorized polyline/polygon. This formulation enables simultaneous optimization through: (1) Differentiable rasterization that enforces pixel-level geometric constraints, and (2) Topology-aware vectorization that maintains structural regularity. Experiments on both nuScenes and Argoverse2 demonstrate that our Gaussian-based representation effectively unifies geometric and topological learning, achieving significant performance improvements and demonstrating strong compatibility with existing HD mapping architectures. Code will be available at https://github.com/peakpang/GSMap
Zhenxuan Zeng, Lingxuan Wang, Sheng Yang +4
School of Computer Science, Northwestern Polytechnical University, China · Unmanned Vehicle Dept, Cainiao Inc., Alibaba Group, China