SMART: Advancing Scalable Map Priors for Driving Topology Reasoning

David Paz; Hengyuan Zhang; Henrik I. Christensen; Junjie Ye; Liu Ren; Xinyu Huang; Yue Wang; Yuliang Guo

REVIEW 2 major objections 6 minor 37 references

SMART: Advancing Scalable Map Priors for Driving Topology Reasoning

T0 review · 2 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A map prior trained from standard-definition and satellite imagery alone can match or beat camera-based lane topology reasoning, and plugging it into online models raises their scores by up to 28 percent on OpenLane-V2.

desk verdict A solid, genuinely novel data recipe for map priors, with one evaluation ambiguity that needs tightening before publication. read the letter →

arxiv 2502.04329 v1 pith:JVTHNQIY submitted 2025-02-06 cs.CV cs.RO

Junjie Ye , David Paz , Hengyuan Zhang , Yuliang Guo , Xinyu Huang , Henrik I. Christensen , Yue Wang , Liu Ren This is my paper · ORCID

classification cs.CVcs.RO

keywords autonomousdrivinglanetopologyreasoningmappriorlearningstandard-definitionmapssatelliteimageryhigh-definitionscalabletrainingOpenLane-V2benchmark

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that lane topology reasoning for autonomous driving can be scaled by replacing sensor-dependent training with a map prior learned from standard-definition (SD) and satellite imagery, supervised by large-scale geo-referenced HD maps. It introduces SMART, a two-stage pipeline: first train a model that predicts lane graphs from SD and satellite maps alone, then freeze that model and inject its features into any camera-based online topology reasoner. Trained on millions of HD map segments from Argoverse 2 motion forecasting, SMART alone outperforms camera-based methods on lane detection and lane-to-lane topology. When plugged into TopoNet and TopoMLP, it improves the OpenLane-V2 score by 20 percent or more, up to 28 percent on the standard split. The paper's broader claim is that geospatial maps plus large HD map corpora can substitute for costly, consistent sensor data.

What carries the argument

The load-bearing object is the fused map-prior feature: SD map polylines are sampled, sinusoidally embedded, and passed through a Transformer encoder, while satellite tiles are encoded with a ResNet backbone; both are sequentially cross-attended into a learnable bird's-eye-view feature map. That BEV representation is decoded into lane instances and a lane-to-lane topological matrix. In the online stage, the same learned BEV query is used either to replace the learnable queries in BEV-based models such as TopoNet or aligned via cross-attention with perspective features in models such as TopoMLP, with SMART's weights frozen to avoid overfitting on small sensor datasets.

What would settle it

Run SMART-OL on a benchmark whose cities were never seen in first-stage training, for example a nuScenes-based evaluation with OSM and Mapbox inputs in Boston or Singapore; if the lane detection and topology improvements over the baseline shrink toward the geo-disjoint numbers or disappear entirely, the claim of a transferable sensor-free map prior is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that a map prior model trained offline on geospatial inputs alone can carry the lane-structure knowledge needed for online driving topology reasoning, without seeing any camera data. SMART fuses SD road polylines and satellite image features into a bird's-eye-view query via cross-attention, then decodes lane centerlines and lane-to-lane connectivity with a deformable-DETR style decoder and a graph network. The result is that the prior alone reaches a lane detection score of 37.9 and a lane topology score of 31.9, above the evaluated online methods, and injecting its features into online models raises lane detection by over 60 percent and lane topology by up to 152 percent while traffic-element detection stays roughly unchanged.

Load-bearing premise

The big gains come from a benchmark in the same cities where the map prior was trained, so the method's usefulness in genuinely new places depends on whether the prior transfers rather than memorizes familiar geometry.

Editorial extensions

If this is right

Sensor data requirements for online topology reasoning can drop substantially: with only 40 percent of sensor data, a SMART-enhanced model matches the full-sensor-data baseline.
The map prior transfers to unseen geographic areas on the paper's geo-disjoint split, improving lane detection by 48.8 percent and lane topology by 148.9 percent over the baseline.
Scaling the prior's training data from 1x to 40x progressively improves both offline and online performance, suggesting that larger HD map corpora directly strengthen the prior.
Fusing SD and satellite modalities is necessary; removing either degrades both offline and online results, with satellite texture contributing more than SD polylines alone.
Traffic-element detection is still done from perspective images, so the prior's benefit is concentrated on lane geometry and lane connectivity rather than on detecting lights and signs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the headline 28 percent gain comes from a split with known geographic overlap between training and evaluation, the 24.9 percent improvement on the geo-disjoint split is the more conservative and arguably the more decisive number for transferability.
If the prior generalizes as claimed, it could be pretrained once over any city with publicly available OSM and Mapbox data and then reused for new cities with no sensor data at all; the paper's scaling experiments point toward this but do not demonstrate it on a truly foreign city.
The paper's remark that priors can be precomputed from a navigation route means the online computational cost could be reduced to near zero, a practical consequence the paper does not quantify.
A direct testable extension would be applying SMART to a benchmark in cities never seen in first-stage training, such as a nuScenes-based evaluation in Boston or Singapore, to confirm the prior transfers across map styles and sensor geometries.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take on the SMART paper. The core idea is sound and the data recipe is genuinely new. Training a lane-topology prior on 4.3M HD maps from Argoverse 2 motion forecasting, using only SD and satellite inputs, is something I haven't seen in the prior work they cite. The two-stage integration—freeze SMART and inject its features into an online mapper—is simple and works. The numbers are strong: SMART-OL lifts TopoNet's OLS by 28% on OpenLane-V2, and the improvement holds across two different online baselines. The scaling study and the SD/satellite ablation are clean and support the main claims.

The soft spot is the geo-disjoint evaluation. They report a 24.9% OLS improvement on a split designed to be unseen, which is the key evidence for transferability. But the paper says they 'resplit the training and validation sets, along with the data used for the first-stage training,' and never explicitly says they retrained SMART on the filtered data. If the frozen SMART was trained on the full 810k maps including areas near the validation locations, then Table IV doesn't actually test transfer—it could be memorization. The small gap between 28.4% and 24.9% makes this concern plausible. The authors need to clarify whether SMART was retrained, and if not, do that experiment. This is a correctable flaw, not a fatal one.

The other overlap issue is the standard split itself. They acknowledge it and cite the leakage papers, so they're aware. But the headline 28% comes from the overlapping split, and the paper doesn't lead with the geo-disjoint result. That's a presentation problem—the clean-split number is the honest one.

Minor points: no code release, no confidence thresholds specified, and traffic-element reasoning is untouched, so the claim about 'driving topology reasoning' is really about lane topology. These are minor.

Overall, this is a serious paper with a real contribution. If the authors clarify the retraining issue, it's a strong conditional accept. I'd send it to a careful referee who knows the OpenLane-V2 leakage literature. The idea will be influential either way, so the evaluation needs to be airtight.

Referee Report

2 major / 6 minor

Summary. The paper introduces SMART, a two-stage pipeline for driving topology reasoning. In the first stage, a map prior model is trained at scale on SD maps and satellite imagery, supervised by HD lane graphs from the Argoverse 2 motion forecasting set (up to roughly 810k HD maps after filtering). In the second stage, this frozen prior is integrated into online topology reasoning baselines (TopoNet and TopoMLP) by replacing BEV queries or via cross-attention, respectively. The paper reports an OpenLane-V2 OLS of 51.1 for SMART-OL(TopoNet), a 28.4% relative improvement over TopoNet, an offline lane-graph performance of DETl 37.9 and TOPll 31.9 from SD/satellite inputs alone, a scaling study showing monotonic gains with more first-stage data, and a geo-disjoint split evaluation reporting a 24.9% OLS improvement in unseen areas.

Significance. If the transferability claim holds, the work is significant: it shows that a map prior can be learned from widely available geospatial data and large-scale geo-referenced HD maps independent of sensor configuration, and that this prior can be injected into online methods with substantial gains. The paper has real strengths: a clean two-stage formulation, thorough ablations for modality fusion and data scaling, a falsifiable scaling trend (Table V), and an explicit attempt to address the known geographic overlap in OpenLane-V2 ([12], [37]). However, the central generalization claim currently rests on a geo-disjoint experiment that is under-specified, which is the main risk to the paper's headline contribution.

major comments (2)

[Section IV-E / Table IV] The manuscript states that 'we resplit the training and validation sets in [8], along with the data used for the first-stage training,' but it never states whether the SMART model used for Table IV was retrained from scratch on the filtered first-stage data. If the same checkpoint trained on all ~810k HD maps (including geo-overlapping ones) is used, the 24.9% OLS improvement could partly reflect memorized map geometry rather than a transferable prior, which would undermine the generalization claim. Please state explicitly whether Table IV uses a retrained SMART; if it does not, the geo-disjoint experiment must be repeated with a SMART retrained on the filtered first-stage data.
[Section IV-E] The geo-disjoint split is described only at a high level. The paper does not report the distance threshold or the method used to partition (e.g., city-level, tile-based, or following the procedure of [37]), nor how the filtering was applied to the Argoverse 2 motion forecasting set. Without this detail, a reader cannot reproduce the split or assess whether the overlap was truly eliminated, which is essential for the paper's central claim of generalization to unseen areas.

minor comments (6)

[Section III-B.3 / Eq. (3)] In Eq. (3), the axis along which the softmax operates is not specified, and the standard 1/sqrt(d) scaling is omitted; please clarify for reproducibility.
[Section III-B.3] There is a typo: 'coodinates' should be 'coordinates.'
[Abstract / Table III] The abstract's 'yielding significant improvements of up to 28%' refers to OLS only; in Table III, DETt for SMART-OL(TopoNet) is -0.6% and for SMART-OL(TopoMLP) is -3.6%. The paper acknowledges this in Remark 3, but the abstract could be more precise by stating that the gain is in overall OLS while traffic-element detection remains roughly unchanged.
[Abstract / Section III-C] The claim that SMART can be 'seamlessly integrated into any online topology reasoning methods' is broader than the evidence, which covers two representative pipelines (BEV-based TopoNet and perspective-based TopoMLP). Suggest tempering 'any' or explicitly noting that broader coverage is future work.
[Section IV-E / Table IV] The baseline in Table IV (OLS 26.9) is substantially lower than the TopoNet baseline in Table III (OLS 39.8). Please clarify that the Table IV baseline is a model retrained on the geo-disjoint reduced training sets and, if so, comment on how the reduced training data affect both baseline and SMART-OL.
[Section IV-F / Table V] The '40x' data amount is approximate; please report the exact number of HD maps for the Base, 18x, and 40x settings for precision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SMART's lane-graph prediction is a standard supervised train/eval setup against external labels; cited self-works are baselines or implementation details, not load-bearing premises.

full rationale

SMART is a two-stage supervised pipeline: a map-prior model is trained on SD/satellite inputs with HD-map lane-graph supervision from Argoverse 2 motion forecasting, and the resulting frozen features are injected into TopoNet/TopoMLP, which are trained and evaluated on OpenLane-V2. The claimed offline performance and the 28% online improvement are measured against external labels under the standard OpenLane-V2 metric; they are not derived from the model's own inputs by construction. The equations (1)-(6) are standard encoders, cross-attention, DETR decoding, GCN, and losses; no parameter is fitted to the reported evaluation and then renamed a prediction. The self-citations in the paper (e.g., TopoOSMR [13] as a baseline, 'following [12]' for sinusoidal embeddings) are not load-bearing: the central claim does not reduce to any theorem or result from the authors' prior work, and no uniqueness/choice-forcing result is invoked. The geographic-overlap concern raised in Sec. IV-E is a data-leakage/soundness issue, not circularity: the paper acknowledges the overlap in OpenLane-V2 and describes a geo-disjoint split, though it does not state explicitly whether the SMART checkpoint in Table IV was retrained on the filtered first-stage data. That ambiguity affects whether the unseen-area numbers are evidence of transfer, but it does not make the headline numbers equivalent to the training objective by definition. Since the evaluation is a standard supervised train/test setup against external benchmarks, no circularity score is warranted.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the availability and representativeness of large-scale HD maps, the alignment of public map sources, and the assumption that lane connectivity is inferable from geospatial inputs. None of these are introduced ad hoc by this paper; they are standard domain assumptions in map learning. The only numerical choice not disclosed is the lane confidence threshold.

free parameters (1)

confidence_threshold = not reported
The lane graph is produced by filtering low-confidence centerlines (Sec. III-B.4), but the threshold value is not stated. It directly determines DETl and TOPll, so its absence prevents exact reproduction.

assumptions (4)

domain assumption HD maps from Argoverse 2 motion forecasting provide a sufficient and representative supervision signal for lane topology learning.
The entire first stage is trained on these maps; if they are not representative of OpenLane-V2 scenes, the prior would not transfer. Invoked in Sec. IV-A.
domain assumption SD maps (OSM) and satellite tiles (Mapbox) are aligned with ego poses and available at the required zoom for all evaluation locations.
The pipeline fetches these at inference; any misalignment or missing tile would degrade the prior. Invoked in Sec. III-B.
domain assumption OpenLane-V2 lane graph labels can be derived from HD maps using the procedure of [8].
The paper states 'Ground-truth lane graphs are derived from these HD maps by regressing centerlines and connecting lanes based on topology [8]' (Sec. IV-A).
domain assumption The lane-to-lane topology is inferable from aerial and SD map inputs.
The offline model predicts TOPll from geospatial inputs alone; this presumes connectivity is visible or inferable from the map, which may fail at complex intersections. Invoked in Sec. III-B.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SMART: Advancing Scalable Map Priors for Driving Topology Reasoning." pith.science (2026). https://pith.science/paper/JVTHNQIY

@misc{pith2026250204329,
  author       = {Pith},
  title        = {Pith review of: SMART: Advancing Scalable Map Priors for Driving Topology Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JVTHNQIY}},
  note         = {Machine review of arXiv:2502.04329}
}

read the original abstract

Topology reasoning is crucial for autonomous driving as it enables comprehensive understanding of connectivity and relationships between lanes and traffic elements. While recent approaches have shown success in perceiving driving topology using vehicle-mounted sensors, their scalability is hindered by the reliance on training data captured by consistent sensor configurations. We identify that the key factor in scalable lane perception and topology reasoning is the elimination of this sensor-dependent feature. To address this, we propose SMART, a scalable solution that leverages easily available standard-definition (SD) and satellite maps to learn a map prior model, supervised by large-scale geo-referenced high-definition (HD) maps independent of sensor settings. Attributed to scaled training, SMART alone achieves superior offline lane topology understanding using only SD and satellite inputs. Extensive experiments further demonstrate that SMART can be seamlessly integrated into any online topology reasoning methods, yielding significant improvements of up to 28% on the OpenLane-V2 benchmark.

Figures

Figures reproduced from arXiv: 2502.04329 by the authors.

**Figure 2.** Outline of the proposed approach. In the first stage (bottom row), SMART is trained at scale using SD and satellite maps for lane graph prediction, supervised by large-scale geo-referenced HD maps. In the second stage (top row), the robust map priors learned by SMART are seamlessly integrated into any online driving topology reasoning models, significantly enhancing lane perception and topology reasoning. 1) SD map … view at source ↗

**Figure 3.** Qualitative comparison of SMART-OL to baselines. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Impact of varying sensor data availability. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 33 canonical work pages

[12]

Augmenting Lane Perception and Topol- ogy Understanding with Standard Definition Navigation Maps,

K. Z. Luo, X. Weng, Y . Wang, S. Wu, J. Li, K. Q. Weinberger, Y . Wang, and M. Pavone, “Augmenting Lane Perception and Topol- ogy Understanding with Standard Definition Navigation Maps,” in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 4029–4035

work page 2024
[37]

Localization Is All You Evaluate: Data Leakage in Online Mapping Datasets and How to Fix It,

A. Lilja, J. Fu, E. Stenborg, and L. Hammarstrand, “Localization Is All You Evaluate: Data Leakage in Online Mapping Datasets and How to Fix It,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2024, pp. 22 150–22 159

work page 2024
[8]

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

H. Wang, T. Li, Y . Li, L. Chen, C. Sima, Z. Liu, B. Wang, P. Jia, Y . Wang, S. Jiang, F. Wen, H. Xu, P. Luo, J. Yan, W. Zhang, and H. Li, “OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping,” in Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , vol. 36, 2023, pp. 18 873–18 884

work page 2023
[1]

HDMapNet: An Online HD Map Construction and Evaluation Framework,

Q. Li, Y . Wang, Y . Wang, and H. Zhao, “HDMapNet: An Online HD Map Construction and Evaluation Framework,” in Proceedings of the International Conference on Robotics and Automation (ICRA) , 2022, pp. 4628–4634

work page 2022
[2]

VectorMapNet: End-to-end Vectorized HD Map Learning,

Y . Liu, T. Yuan, Y . Wang, Y . Wang, and H. Zhao, “VectorMapNet: End-to-end Vectorized HD Map Learning,” in Proceedings of the International Conference on Machine Learning (ICML) , vol. 202, 2023, pp. 22 352–22 369

work page 2023
[3]

MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction,

B. Liao, S. Chen, X. Wang, T. Cheng, Q. Zhang, W. Liu, and C. Huang, “MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction,” in Proceedings of the International Conference on Learning Representations (ICLR) , 2023, pp. 1–18

work page 2023
[4]

MapTRv2: An End-to-End Framework for Online Vectorized HD Map Construction,

B. Liao, S. Chen, Y . Zhang, B. Jiang, Q. Zhang, W. Liu, C. Huang, and X. Wang, “MapTRv2: An End-to-End Framework for Online Vectorized HD Map Construction,” International Journal of Computer Vision, pp. 1–17, 2024

work page 2024
[5]

Perceive, Predict, and Plan: Safe Motion Planning Through Inter- pretable Semantic Representations,

A. Sadat, S. Casas, M. Ren, X. Wu, P. Dhawan, and R. Urtasun, “Perceive, Predict, and Plan: Safe Motion Planning Through Inter- pretable Semantic Representations,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2020, pp. 414–430

work page 2020

Show all 37 references

[6]

Wayformer: Motion Forecasting via Simple & Efficient Attention Networks,

N. Nayakanti, R. Al-Rfou, A. Zhou, K. Goel, K. S. Refaat, and B. Sapp, “Wayformer: Motion Forecasting via Simple & Efficient Attention Networks,” in Proceedings of the IEEE International Con- ference on Robotics and Automation (ICRA) , 2023, pp. 2980–2987

2023
[7]

A Language Agent for Autonomous Driving,

J. Mao, J. Ye, Y . Qian, M. Pavone, and Y . Wang, “A Language Agent for Autonomous Driving,” in Proceedings of the Conference on Language Modeling (COLM) , 2024, pp. 1–40

2024
[9]

Graph-based Topology Reasoning for Driving Scenes,

T. Li, L. Chen, H. Wang, Y . Li, J. Yang, X. Geng, S. Jiang, Y . Wang, H. Xu, C. Xu, J. Yan, P. Luo, and H. Li, “Graph-based Topology Reasoning for Driving Scenes,” arXiv preprint arXiv:2304.05277, pp. 1–12, 2023

2023 arXiv
[10]

TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning,

D. Wu, J. Chang, F. Jia, Y . Liu, T. Wang, and J. Shen, “TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning,” in Pro- ceedings of the International Conference on Learning Representations (ICLR), 2024, pp. 1–12

2024
[11]

RoadPainter: Points Are Ideal Navigators for Topology transformER,

Z. Ma, S. Liang, Y . Wen, W. Lu, and G. Wan, “RoadPainter: Points Are Ideal Navigators for Topology transformER,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2024, pp. 1– 17

2024
[13]

Enhancing Online Road Network Perception and Reasoning with Standard Definition Maps,

H. Zhang, D. Paz, Y . Guo, A. Das, X. Huang, K. Haug, H. I. Christensen, and L. Ren, “Enhancing Online Road Network Perception and Reasoning with Standard Definition Maps,” in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 1–9

2024
[14]

Argoverse 2: Next Generation Datasets for Self- Driving Perception and Forecasting,

B. Wilson, W. Qi, T. Agarwal, J. Lambert, J. Singh, S. Khandelwal, B. Pan, R. Kumar, A. Hartnett, J. Kaesemodel Pontes, D. Ramanan, P. Carr, and J. Hays, “Argoverse 2: Next Generation Datasets for Self- Driving Perception and Forecasting,” in Proceedings of the Neural Informat...

2021
[15]

NuPlan: A closed-loop ML- based planning benchmark for autonomous vehicles,

H. Caesar, J. Kabzan, K. S. Tan, W. K. Fong, E. Wolff, A. Lang, L. Fletcher, O. Beijbom, and S. Omari, “NuPlan: A closed-loop ML- based planning benchmark for autonomous vehicles,” arXiv preprint arXiv:2106.11810, pp. 1–5, 2021

2021 arXiv
[16]

Scalability in Perception for Autonomous Driving: Waymo Open Dataset,

P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V . Patnaik, P. Tsui, J. Guo, Y . Zhou, Y . Chai, B. Caine, V . Vasudevan, W. Han, J. Ngiam, H. Zhao, A. Timofeev, S. Ettinger, M. Krivokon, A. Gao, A. Joshi, Y . Zhang, J. Shlens, Z. Chen, and D. Anguelov, “Scalability in Perc...

2020
[17]

Openstreetmap,

“Openstreetmap,” https://www.openstreetmap.org, 2024, accessed: 2024-09-13

2024
[18]

Mapbox Raster Tiles API,

“Mapbox Raster Tiles API,” https://docs.mapbox.com/api/maps/ raster-tiles/, 2025, accessed: 2025-01-29

2025
[19]

AutoGraph: Predicting Lane Graphs From Traffic Observations,

J. Z ¨urn, I. Posner, and W. Burgard, “AutoGraph: Predicting Lane Graphs From Traffic Observations,” IEEE Robotics and Automation Letters, vol. 9, no. 1, pp. 73–80, 2024

2024
[20]

Learning and Aggregating Lane Graphs for Urban Automated Driv- ing,

M. B ¨uchner, J. Z ¨urn, I.-G. Todoran, A. Valada, and W. Burgard, “Learning and Aggregating Lane Graphs for Urban Automated Driv- ing,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2023, pp. 13 415–13 424

2023
[21]

Bezier Everywhere All at Once: Learning Drivable Lanes as Bezier Graphs,

H. Blayney, H. Tian, H. Scott, N. Goldbeck, C. Stetson, and P. An- geloudis, “Bezier Everywhere All at Once: Learning Drivable Lanes as Bezier Graphs,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2024, pp. 15 365– 15 374

2024
[22]

TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes,

Y . Fu, W. Liao, X. Liu, H. xu, Y . Ma, F. Dai, and Y . Zhang, “TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes,” arXiv preprint arXiv:2405.14747, pp. 1–12, 2024

2024 arXiv
[23]

Enhancing 3D Lane Detection and Topology Reasoning with 2D Lane Priors,

H. Li, Z. Huang, Z. Wang, W. Rong, N. Wang, and S. Liu, “Enhancing 3D Lane Detection and Topology Reasoning with 2D Lane Priors,” arXiv preprint arXiv:2406.03105 , pp. 1–20, 2024

2024 arXiv
[24]

Neural Map Prior for Autonomous Driving,

X. Xiong, Y . Liu, T. Yuan, Y . Wang, Y . Wang, and H. Zhao, “Neural Map Prior for Autonomous Driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2023, pp. 17 535–17 544

2023
[25]

Comple- menting Onboard Sensors with Satellite Maps: A New Perspective for HD Map Construction,

W. Gao, J. Fu, Y . Shen, H. Jing, S. Chen, and N. Zheng, “Comple- menting Onboard Sensors with Satellite Maps: A New Perspective for HD Map Construction,” in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , 2024, pp. 11 103– 11 109

2024
[26]

P-MapNet: Far-Seeing Map Generator Enhanced by Both SDMap and HDMap Priors,

Z. Jiang, Z. Zhu, P. Li, H.-a. Gao, T. Yuan, Y . Shi, H. Zhao, and H. Zhao, “P-MapNet: Far-Seeing Map Generator Enhanced by Both SDMap and HDMap Priors,” IEEE Robotics and Automation Letters , vol. 9, no. 10, pp. 8539–8546, 2024

2024
[27]

PreSight: Enhancing Autonomous Vehicle Perception with City-Scale NeRF Priors,

T. Yuan, Y . Mao, J. Yang, Y . Liu, Y . Wang, and H. Zhao, “PreSight: Enhancing Autonomous Vehicle Perception with City-Scale NeRF Priors,” in Proceedings of the European Conference on Computer Vision (ECCV), 2024, pp. 323–339

2024
[28]

Attention Is All you Need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention Is All you Need,” in Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), vol. 30. Curran Associates, Inc., 2017, pp. 1–11

2017
[29]

Deformable DETR: Deformable Transformers for End-to-End Object Detection,

X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable DETR: Deformable Transformers for End-to-End Object Detection,” in Pro- ceedings of the International Conference on Learning Representations (ICLR), 2021, pp. 1–16

2021
[30]

BEVFormer: Learning Bird’s-Eye-View Representation from Multi- Camera Images via Spatiotemporal Transformers,

Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Y . Qiao, and J. Dai, “BEVFormer: Learning Bird’s-Eye-View Representation from Multi- Camera Images via Spatiotemporal Transformers,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2022, pp. 1– 18

2022
[31]

nuScenes: A Multimodal Dataset for Autonomous Driving,

H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuScenes: A Multimodal Dataset for Autonomous Driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 1...

2020
[32]

Structured Bird’s-Eye-View Traffic Scene Understanding from Onboard Images,

Y . B. Can, A. Liniger, D. P. Paudel, and L. Van Gool, “Structured Bird’s-Eye-View Traffic Scene Understanding from Onboard Images,” in Proceedings of the IEEE/CVF International Conference on Com- puter Vision (ICCV) , 2021, pp. 15 641–15 650

2021
[33]

Computing discrete Fr ´echet distance,

T. Eiter and H. Mannila, “Computing discrete Fr ´echet distance,” 1994

1994
[34]

Deep Residual Learning for Image Recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

2016
[35]

ImageNet Classifi- cation with Deep Convolutional Neural Networks,

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classifi- cation with Deep Convolutional Neural Networks,” in Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , F. Pereira, C. Burges, L. Bottou, and K. Weinberger, Eds., vol. 25, 2012, pp. 1–9

2012
[36]

Decoupled Weight Decay Regulariza- tion,

I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regulariza- tion,” in Proceedings of the International Conference on Learning Representations (ICLR), 2019, pp. 1–19

2019

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [12]

Augmenting Lane Perception and Topol- ogy Understanding with Standard Definition Navigation Maps,

K. Z. Luo, X. Weng, Y . Wang, S. Wu, J. Li, K. Q. Weinberger, Y . Wang, and M. Pavone, “Augmenting Lane Perception and Topol- ogy Understanding with Standard Definition Navigation Maps,” in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 4029–4035

work page 2024

[2] [37]

Localization Is All You Evaluate: Data Leakage in Online Mapping Datasets and How to Fix It,

A. Lilja, J. Fu, E. Stenborg, and L. Hammarstrand, “Localization Is All You Evaluate: Data Leakage in Online Mapping Datasets and How to Fix It,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2024, pp. 22 150–22 159

work page 2024

[3] [8]

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

H. Wang, T. Li, Y . Li, L. Chen, C. Sima, Z. Liu, B. Wang, P. Jia, Y . Wang, S. Jiang, F. Wen, H. Xu, P. Luo, J. Yan, W. Zhang, and H. Li, “OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping,” in Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , vol. 36, 2023, pp. 18 873–18 884

work page 2023

[4] [1]

HDMapNet: An Online HD Map Construction and Evaluation Framework,

Q. Li, Y . Wang, Y . Wang, and H. Zhao, “HDMapNet: An Online HD Map Construction and Evaluation Framework,” in Proceedings of the International Conference on Robotics and Automation (ICRA) , 2022, pp. 4628–4634

work page 2022

[5] [2]

VectorMapNet: End-to-end Vectorized HD Map Learning,

Y . Liu, T. Yuan, Y . Wang, Y . Wang, and H. Zhao, “VectorMapNet: End-to-end Vectorized HD Map Learning,” in Proceedings of the International Conference on Machine Learning (ICML) , vol. 202, 2023, pp. 22 352–22 369

work page 2023

[6] [3]

MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction,

B. Liao, S. Chen, X. Wang, T. Cheng, Q. Zhang, W. Liu, and C. Huang, “MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction,” in Proceedings of the International Conference on Learning Representations (ICLR) , 2023, pp. 1–18

work page 2023

[7] [4]

MapTRv2: An End-to-End Framework for Online Vectorized HD Map Construction,

B. Liao, S. Chen, Y . Zhang, B. Jiang, Q. Zhang, W. Liu, C. Huang, and X. Wang, “MapTRv2: An End-to-End Framework for Online Vectorized HD Map Construction,” International Journal of Computer Vision, pp. 1–17, 2024

work page 2024

[8] [5]

Perceive, Predict, and Plan: Safe Motion Planning Through Inter- pretable Semantic Representations,

A. Sadat, S. Casas, M. Ren, X. Wu, P. Dhawan, and R. Urtasun, “Perceive, Predict, and Plan: Safe Motion Planning Through Inter- pretable Semantic Representations,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2020, pp. 414–430

work page 2020