Pith. sign in

REVIEW 4 major objections 6 minor 27 references

Using Language and Road Manuals to Inform Map Reconstruction for Autonomous Driving

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding OpenStreetMap metadata embeddings and design-manual lane-width priors to SMERF improves all evaluated lane-topology metrics, raising OpenLane-V2 Score from 0.5894 to 0.5995.

desk verdict A plausible lightweight augmentation to SMERF that is undermined by an evaluation with no held-out split or error bars on 64 frames. read the letter →

arxiv 2506.10317 v2 pith:3LSQAYMM submitted 2025-06-12 cs.RO cs.AI

Akshar Tumu , Henrik I. Christensen , Marcell Vazquez-Chanlatte , Chikao Tsuchiya , Dhaval Bhanderi This is my paper · ORCID
classification cs.ROcs.AI
keywords lanetopologypredictionautonomousdrivingstandarddefinitionmapsOpenStreetMapmetadataretrievalaugmentedgenerationwidthpriorsSMERFLane-V2
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

The paper is trying to establish that language-derived priors—road metadata from OpenStreetMap and lane-width values extracted from a highway design manual—can improve online lane-topology prediction for autonomous driving. It adds text embeddings of these priors to SMERF, a standard-definition-map-based model, by mapping the text embeddings into the same space as the road-polyline embeddings and adding them before cross-attention with camera features. On two complex intersection scenarios from OpenLane-V2, the best configuration improves the OpenLane-V2 Score from 0.5894 to 0.5995, with lane-lane connectivity accuracy up 16% relative to baseline. If the claim holds, richer use of widely available semantic map data can narrow the gap to high-definition maps at a fraction of the maintenance cost.

What carries the argument

The carrying mechanism is a fused text-plus-graph embedding per road polyline. For each road, the paper concatenates OpenStreetMap metadata fields (road suffix, service type, oneway, number of lanes, maxspeed, and similar tags) into a single string, embeds it with a text-embedding model, and separately obtains a lane-width embedding by querying a retrieval-augmented LLM over a highway design manual. The two text embeddings are summed as $t_p = o_p + l_p$, passed through a 2-layer MLP to align with SMERF's map-embedding space, and added to the polyline graph embedding: $e_p = G(p) + \mathrm{MLP}(t_p)$. This combined embedding is fed into SMERF's BEV constructor for cross-attention with multi-camera BEV features, so the language priors influence the decoder without changing the model architecture.

What would settle it

Measure lane widths on the two test intersections and retrain the best configuration with the measured values replacing the retrieval-extracted widths; if TOP_ll does not improve over baseline, the design-manual prior is not the source of the gain, and if it improves more, the retrieved values were actually degrading the prior.

Watch

Extended reading notes

Core claim

The central claim is that augmenting SMERF's road-centerline encodings with text embeddings of OpenStreetMap metadata and retrieval-augmented lane-width priors yields consistent gains across all aspects of lane-topology prediction—lane detection, traffic-element detection, lane-lane connectivity, and lane-traffic-element association—on the two evaluated complex intersections. Concretely, the paper reports that the best configuration (NF + λ) improves the OpenLane-V2 Score from 0.5894 to 0.5995 and TOP_ll by 16% relative to the SMERF baseline, while the F1 + RAG configuration gives the best lane-centerline detection. The authors take this as evidence that natural-language conventions embedded in road names and design manuals carry reusable geometric and functional cues that generalize across geographies without high-definition maps.

Load-bearing premise

The lane-width priors taken from the California design manual are assumed to match actual lane geometry in Detroit and Pittsburgh, though the paper reports no check against measured widths.

Editorial extensions

If this is right

  • The best configuration (NF + λ) raises the OpenLane-V2 Score from 0.5894 to 0.5995 on the two test scenarios, with lane-lane connectivity (TOP_ll) improving by 16% relative to baseline.
  • Including OpenStreetMap metadata and lane-width embeddings requires only a 2-layer MLP plus text embeddings, leaving SMERF's architecture unchanged and adding no high-definition-map dependency.
  • Improvements concentrate in topology association metrics (TOP_ll and TOP_lt) rather than raw detection, consistent with the claim that semantic road priors mainly disambiguate connectivity and lane-to-traffic-element links.
  • Configurations that freeze most of the network (F0) underperform, while unfreezing (NF) gives the best results, indicating the text prior needs the perception stack to adapt to it.
  • Across the two geographically distinct scenarios, the method improves all four topology-aware metrics for the best configuration, supporting the paper's claim of geographical generalization.

Reading between the lines

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

  • Beyond the paper: the same fusion rule could be applied to other standard-definition-map-conditioned online mapping models, since it changes only how the map embedding is formed and leaves the decoder untouched.
  • Beyond the paper: the road suffix alone (Avenue, Street, Circle, Expressway) might provide a usable prior in areas where other OpenStreetMap metadata tags are sparse, because the paper already extracts and embeds the suffix while excluding the full road name.
  • Beyond the paper: querying the retrieval pipeline with each city's own design manual rather than a single state manual would separate genuine geographic transfer from prompt-reasoning gains; the two test cities lie outside the manual's jurisdiction.
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.

Referee Report

4 major / 6 minor

Summary. The paper proposes augmenting SMERF, an online lane-topology prediction model, with (1) embeddings of OpenStreetMap road metadata (beyond the geometry and highway type already used) and (2) lane-width priors extracted from a road design manual via retrieval-augmented generation (RAG). The text embeddings are transformed by a two-layer MLP and added to SMERF's graph embeddings, optionally with a learnable fusion weight λ (Eq. 2). The authors evaluate nine training/fusion configurations on two OpenLane-V2 scenarios totaling 64 frames and report that the best configuration ('NF + λ') improves the OpenLane-V2 Score from 0.5894 to 0.5995, with a 16% relative improvement in TOP_ll. They conclude that language and manual-derived priors yield consistent improvement in all aspects of lane-topology prediction and support generalization to diverse geographies.

Significance. The core idea—using language-encoded road semantics and design-manual priors as lightweight additions to an SD-map-based mapping model—is interesting and potentially useful for scaling to geographies without HD maps. The fusion is parameter-efficient and the paper provides a clear table of nine configurations. However, the experimental support is currently insufficient to establish generalization: the evaluation uses only 64 frames from two scenarios for both training and evaluation, with no held-out split, no error bars, and no multiple-seed analysis; the best configuration is selected after comparing nine variants on the same data. The geographic transferability of California design-manual lane widths to Detroit and Pittsburgh is asserted but not validated. If a rigorous evaluation protocol is provided, the method could be a valuable contribution, but as presented the central claims are not supported.

major comments (4)
  1. [Section IV-A and Section V] The evaluation protocol does not separate training and evaluation frames. Section IV-A states that the two scenarios 'have 64 frames in total' and are selected 'for training and evaluating our model,' and Section IV-C describes a second round of 100 epochs of training on the model configurations, while Section V reports all metrics 'over the two selected scenarios.' Without a held-out split, the reported improvements (e.g., TOP_ll from 0.0595 to 0.0674 in Table I) could reflect memorization of the evaluation geometry rather than generalization. Additionally, the best configuration is selected after comparing nine variants on the same 64 frames, which introduces selection bias. The paper should provide a proper train/validation split, per-scenario results, and repeated runs with multiple seeds to report means and variances.
  2. [Section III-B and Table I] The RAG pipeline extracts lane-width priors from the California Highway Design Manual and applies them to roads in Detroit, Michigan and Pittsburgh, Pennsylvania without any geographic adaptation or validation. If the manual's standard lane widths do not match the actual lane geometry in those cities, the injected priors are systematically wrong, which could either degrade performance or artificially inflate the reported gains on the two test sites. The authors should validate the RAG-extracted lane widths against the ground-truth lane widths in the evaluation scenarios, or explicitly justify why California design-manual widths transfer to these locations.
  3. [Section VII and Table I] The claim of 'consistent improvement in all aspects of lane-topology prediction for the chosen complex scenarios' is not supported by Table I. The table shows that F0 variants degrade substantially on all four metrics relative to the baseline, F1+λ degrades TOP_lt relative to F1 (0.3462 vs. 0.3519), and NF degrades DET_l and DET_t relative to the baseline. Only the best configuration shows improvement on a subset of metrics, and the paper's own text in Section V notes that the NF+λ variant performs best on three of the five metrics. The conclusion should be limited to the best configuration and acknowledge the variability across configurations.
  4. [Table I] The reported gains lack statistical support. The TOP_ll improvement from 0.0595 to 0.0674 is a relative increase of about 16%, but the absolute difference is 0.008, and with only 64 frames and no error bars or significance tests it is unclear whether this difference is meaningful. The authors should report variance across runs or perform a significance test to establish that the difference is not due to noise.
minor comments (6)
  1. [Abstract and Section I] The abbreviation 'RAG' is used in the abstract and Section I but is not spelled out as 'Retrieval Augmented Generation' until Section III-B; please define it at first use.
  2. [Section III-B] Reference [17] (Touvron et al.) is cited for 'Llama3.3 LLM,' but [17] is the original LLaMA paper; please update the citation to the specific Llama 3.3 release or correct the model name.
  3. [Section III-A] The text says 'OpenAI's text-embedding-3-small model' without a version or access date; please include the model version and retrieval date for reproducibility.
  4. [Section V] The phrase 'Among all four metrics' is ambiguous because Section IV-B lists five metrics (DET_l, DET_t, TOP_ll, TOP_lt, and OLS); please clarify that the comparison refers to the four topology-aware metrics excluding OLS.
  5. [Table I] The learned λ value is reported as 0.869 for all three configurations that use it; please explain why the learned value is identical across configurations and report the initialization and training details for λ.
  6. [Figure 2] The caption states that lane colors indicate intended actions at intersections, but the figure appears to be rendered in grayscale in the manuscript; please use a color version or annotate the figure to convey this information.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported gains are measured on the same 64 frames used for training, so the Table I 'predictions' are in-sample fits rather than generalization evidence.

  1. fitted input called prediction [Section IV-A (Dataset), IV-C (Training Configurations), V (Results/Table I); Section VI generalization claim]
    "We select two complex road scenarios from the OpenLane-V2 dataset, which are from different cities in the United States, for training and evaluating our model. The two scenarios have 64 frames in total. ... We first start by training the original SMERF model for 100 epochs. Then we further train the model for a second round of 100 epochs by including the OSM Metadata embeddings and Lane-width embeddings. ... Table I shows the results over the two selected scenarios for the baseline (SMERF) and our different model configurations."

    The 64 frames named in Section IV-A as the training data are exactly the frames on which Table I reports DET_l, DET_t, TOP_ll, TOP_lt, and OLS. The model weights—including the added 2-layer MLP and the learnable fusion weight λ—are fit to those frames (Section IV-C), so every metric in Table I measures reconstruction of the training set, not prediction. The claimed +16% TOP_ll and OLS gain 0.5894→0.5995 are therefore in-sample fits presented as evidence of generalization, and the Section VI statement that the setup 'confirms the generalization ability of our method' is unsupported by any held-out measurement. No train/validation split, per-scenario breakdown, or seed variance is reported; selecting the best of nine configurations on the same 64 frames compounds the selection bias.

full rationale

The paper's input-side derivation is not circular: OSM metadata and lane-width priors from the Caltrans Highway Design Manual are external, RAG-extracted content added to SMERF via Eqs. (1)-(3), and no equation defines the predicted lane topology as equal to those embeddings; the fusion weight λ is learned, not fixed. Self-citations to the authors' prior SD++ paper provide the RAG pipeline and a motivating result, but they are not load-bearing for the present empirical claim, which is tested (however imperfectly) against a baseline. The substantive circularity is in the evaluation protocol: Section IV-A selects two scenarios 'for training and evaluating our model,' and Section V reports all metrics 'over the two selected scenarios.' Since training runs for 200 epochs on those same 64 frames (Section IV-C), Table I's numbers are training-set fits. The paper then uses these in-sample numbers to assert generalization ('confirms the generalization ability of our method,' Section VI; 'consistent improvement in all aspects,' Section VII). That is a fitted input called prediction: the reported improvements reduce to how well the model memorized the evaluation frames, not to any out-of-sample benefit of the language/road-manual priors. Score 6 reflects this central evaluation circularity; the method itself is not defined in terms of its outputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical or abstract entities are postulated; the method uses existing OSM metadata and road design manuals. The only additions are a learned MLP and a fusion weight, which are standard model components. The main hidden assumptions are geographic transferability of the California design manual, completeness of OSM metadata, and reliability of commercial LLM embeddings.

free parameters (2)
  • learnable fusion weight lambda = 0.869
    A scalar weight in Eq. 2 controlling the contribution of text embeddings to the fused map embedding, learned during the second round of training and reported as 0.869 in Table I.
  • second-round training epochs = 100
    The model is trained for an additional 100 epochs with text embeddings; the schedule is chosen by the authors, and the paper does not report learning rate, batch size, or other hyperparameters.
assumptions (4)
  • domain assumption OSM metadata fields (service, oneway, lanes, maxspeed, road suffix) are informative for lane topology and are present and accurate for the evaluated roads.
    The method's central premise, introduced in Section III-A and used throughout the fusion and evaluation.
  • domain assumption The California Highway Design Manual provides lane widths that are valid for roads in Detroit, MI and Pittsburgh, PA, and the RAG pipeline extracts them correctly.
    Section III-B uses the Caltrans manual via RAG without geographic adaptation or validation of the extracted widths.
  • domain assumption OpenAI text-embedding-3-small and Llama3.3 embeddings are stable and meaningful representations for this task.
    These model outputs are used without reproducibility or stability checks in Sections III-A and III-B.
  • domain assumption SMERF is a reasonable baseline for online lane-topology prediction with SD-map priors.
    The paper builds on SMERF [13] and does not independently verify the baseline's behavior beyond the reported numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Using Language and Road Manuals to Inform Map Reconstruction for Autonomous Driving." pith.science (2026). https://pith.science/paper/3LSQAYMM

@misc{pith2026250610317,
  author       = {Pith},
  title        = {Pith review of: Using Language and Road Manuals to Inform Map Reconstruction for Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3LSQAYMM}},
  note         = {Machine review of arXiv:2506.10317}
}
read the original abstract

Lane-topology prediction is a critical component of safe and reliable autonomous navigation. An accurate understanding of the road environment aids this task. We observe that this information often follows conventions encoded in natural language, through design codes that reflect the road structure and road names that capture the road functionality. We augment this information in a lightweight manner to SMERF, a map-prior-based online lane-topology prediction model, by combining structured road metadata from OSM maps and lane-width priors from Road design manuals with the road centerline encodings. We evaluate our method on two geo-diverse complex intersection scenarios. Our method shows improvement in both lane and traffic element detection and their association. We report results using four topology-aware metrics to comprehensively assess the model performance. These results demonstrate the ability of our approach to generalize and scale to diverse topologies and conditions.

Figures

Figures reproduced from arXiv: 2506.10317 by the authors.

Figure 1
Figure 1. RAG Prompt [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the front-view image and corresponding Ground-Truth BEV lane-topologies. Lane colors in the BEV Map indicate intended actions [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Lane-topology predictions in the BEV view for the two scenarios. The [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 15 canonical work pages

  1. [1]

    Packt Publishing Ltd,

    Jonathan Bennett.OpenStreetMap. Packt Publishing Ltd,

    work page

  2. [2]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. URL https://openaccess.thecvf.com/content CVPR 202...

    work page 2020

  3. [3]

    Alejandro Diaz-Diaz, Manuel Oca ˜na, ´Angel Llamazares, Carlos G ´omez-Hu´elamo, Pedro Revenga, and Luis M. Bergasa. HD maps: Exploiting OpenDRIVE potential for Path Planning and Map Monitoring. In2022 IEEE Intelligent Vehicles Symposium (IV), page 1211–1217. IEEE Press, 2022. doi: 10.1109/IV51971.2022.9827297. URL https://doi.org/10.1109/IV51971.2022.9827297

    arXiv 2022

  4. [4]

    Christensen, Marcell Vazquez-Chanlatte, and Chikao Tsuchiya

    Hitvarth Diwanji, Jing-Yan Liao, Akshar Tumu, Henrik I. Christensen, Marcell Vazquez-Chanlatte, and Chikao Tsuchiya. SD++: Enhancing Standard Definition Maps by Incorporating Road Knowledge using LLMs. In Intelligent Vehicles - 2025, June 2025. URL https: //arxiv.org/abs/2502.02773

    arXiv 2025

  5. [5]

    DenseTNT: End- to-End Trajectory Prediction From Dense Goal Sets

    Junru Gu, Chen Sun, and Hang Zhao. DenseTNT: End- to-End Trajectory Prediction From Dense Goal Sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15303–15312, Octo- ber 2021

    work page 2021

  6. [6]

    Producing and leveraging online map uncertainty in trajectory prediction

    Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14521–14530, June 2024. URL https: //openaccess.thecvf.com/content/CVPR2024/html/Gu Producing and Leverag...

    work page 2024

  7. [7]

    Accelerating online mapping and behavior prediction via direct bev feature atten- tion

    Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Accelerating online mapping and behavior prediction via direct bev feature atten- tion. InComputer Vision – ECCV 2024: 18th Euro- pean Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part LXXXI, page 412–428, Berlin, Heidelberg, 2024. Springer-Verlag. ISBN...

    doi:10.1007/978-3-031-73004-7 2024

  8. [8]

    HDM-RRT: A Fast HD-Map- Guided Motion Planning Algorithm for Autonomous Driving in the Campus Environment.Remote Sensing, 15(2), 2023

    Xiaomin Guo, Yongxing Cao, Jian Zhou, Yuanxian Huang, and Bijun Li. HDM-RRT: A Fast HD-Map- Guided Motion Planning Algorithm for Autonomous Driving in the Campus Environment.Remote Sensing, 15(2), 2023. ISSN 2072-4292. doi: 10.3390/rs15020487. URL https://www.mdpi.com/2072-4292/15/2/487

    work page doi:10.3390/rs15020487 2023

Show all 27 references
  1. [9]

    ImagineMap: En- hanced HD Map Construction with SD Maps, 2024

    Yishen Ji, Zhiqi Li, and Tong Lu. ImagineMap: En- hanced HD Map Construction with SD Maps, 2024. URL https://arxiv.org/abs/2412.16938

    2024 arXiv

  2. [10]

    HDMap- Net: An Online HD Map Construction and Evalua- tion Framework

    Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. HDMap- Net: An Online HD Map Construction and Evalua- tion Framework. In2022 International Conference on Robotics and Automation (ICRA), pages 4628–4634,

  3. [11]

    Maptrv2: An end-to-end framework for online vectorized hd map construction.International Journal of Computer Vision, pages 1–23, 2024

    Bencheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction.International Journal of Computer Vision, pages 1–23, 2024. URL https://doi. org/10.1007/s11263...

    2024 doi

  4. [12]

    VectorMapNet: end-to-end vectorized HD map learning

    Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. VectorMapNet: end-to-end vectorized HD map learning. InProceedings of the 40th Inter- national Conference on Machine Learning, ICML’23. JMLR.org, 2023. URL https://proceedings.mlr.press/ v202/liu23ax.html

    2023

  5. [13]

    Augmenting Lane Perception and Topology Un- derstanding with Standard Definition Navigation Maps

    Katie Z Luo, Xinshuo Weng, Yan Wang, Shuang Wu, Jie Li, Kilian Q Weinberger, Yue Wang, and Marco Pavone. Augmenting Lane Perception and Topology Un- derstanding with Standard Definition Navigation Maps. In2024 IEEE International Conference on Robotics and Automation (ICRA), pa...

    2024

  6. [14]

    Department of Transportation

    California. Department of Transportation. California Department of Transportation (Caltrans): Highway De- sign Manual: U.S. Customary Units: Seventh Edition. https://rosap.ntl.bts.gov/view/dot/15907, 2020

    2020

  7. [15]

    Uni-PrevPredMap: Extending Pre- vPredMap to a Unified Framework of Prior-Informed Modeling for Online Vectorized HD Map Construction,

    Nan Peng, Xun Zhou, Mingming Wang, Guisong Chen, and Wenqi Xu. Uni-PrevPredMap: Extending Pre- vPredMap to a Unified Framework of Prior-Informed Modeling for Online Vectorized HD Map Construction,

  8. [16]

    Scalability in Perception for Autonomous Driving: Waymo Open Dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang,...

    2020

  9. [17]

    LLaMA: Open and Efficient Foundation Language Models, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth ´ee Lacroix, Bap- tiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and Efficient Foundation Lan...

    2023 arXiv

  10. [18]

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

    Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping. In A. Oh, T. ...

    2023

  11. [19]

    Argoverse 2: Next Generation Datasets for Self- Driving Perception and Forecasting, 2023

    Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaese- model Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next Generation Datasets for Self- Driving Perception and...

    2023 arXiv

  12. [20]

    Neural Map Prior for Autonomous Driving

    Xuan Xiong, Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Neural Map Prior for Autonomous Driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17535–17544, June 2023. URL https://openaccess.thecvf.com/con...

    2023

  13. [21]

    TopoSD: Topology-Enhanced Lane Segment Per- ception with SDMap Prior, 2024

    Sen Yang, Minyue Jiang, Ziwei Fan, Xiaolu Xie, Xiao Tan, Yingying Li, Errui Ding, Liang Wang, and Jingdong Wang. TopoSD: Topology-Enhanced Lane Segment Per- ception with SDMap Prior, 2024. URL https://arxiv.org/ abs/2411.14751

    2024 arXiv

  14. [22]

    StreamMapNet: Streaming Mapping Network for Vectorized Online HD Map Construction

    Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. StreamMapNet: Streaming Mapping Network for Vectorized Online HD Map Construction. InWACV, pages 7341–7350, 2024. URL https://doi.org/ 10.1109/W ACV57701.2024.00719

    2024

  15. [23]

    HiVT: Hierarchical Vector Transformer for Multi-Agent Motion Prediction

    Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. HiVT: Hierarchical Vector Transformer for Multi-Agent Motion Prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8823–8833, June 2022. URL https://openaccess.t...

    2022

  16. [24]

    Deformable detr: Deformable transformers for end-to-end object detection, 2021

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection, 2021. URL https://arxiv.org/abs/2010.04159

    2021 arXiv

  17. [2010]

    URL https://www.researchgate.net/publication/ 262206447 OpenStreetMap

  18. [2022]

    URL https://doi.org/10.1109/ICRA46639.2022.9812383

    doi: 10.1109/ICRA46639.2022.9812383. URL https://doi.org/10.1109/ICRA46639.2022.9812383

    2022

  19. [2025]

    URL https://arxiv.org/abs/2504.06647

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.