pith. sign in

arxiv: 2603.10688 · v2 · pith:N6UTBGGTnew · submitted 2026-03-11 · 💻 cs.RO · cs.CV

MapGCLR: Geospatial Contrastive Learning of Representations for Online Vectorized HD Map Construction

Jonas Merkert , Alexander Blumberg , Jan-Hendrik Pauls , Christoph Stiller This is my paper

Pith reviewed 2026-05-25 06:42 UTC · model grok-4.3

classification 💻 cs.RO cs.CV
keywords online HD map constructioncontrastive learningBEV feature gridsgeospatial consistencysemi-supervised learningvectorized mapsautonomous vehiclesself-supervised representation learning
0
0 comments X

The pith

Enforcing geospatial consistency between overlapping BEV grids via contrastive loss improves vectorized online HD map construction over supervised baselines.

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

The paper seeks to establish that a contrastive loss enforcing geospatial consistency on pairs of birds-eye-view feature grids from overlapping traversals can strengthen the latent representation inside a vectorized HD map model. This is achieved by first analyzing dataset overlaps to produce multi-traversal splits, then training the model in a semi-supervised regime: supervised on a small set of single-traversal labels and self-supervised on a larger set of unlabeled multi-traversal data. A sympathetic reader would care because the method promises to lower the volume of expensive map annotations required while raising downstream map-construction accuracy.

Core claim

By generating subsidiary dataset splits that satisfy adjustable multi-traversal overlap requirements and then applying a contrastive loss that pulls together BEV feature grids whose geospatial footprints overlap, the same model architecture reaches higher vectorized map perception scores than a purely supervised baseline trained only on reduced single-traversal labels; the improvement is visible both in quantitative map metrics and in clearer class separation within PCA projections of the BEV feature space.

What carries the argument

The overlap-analysis procedure that produces multi-traversal dataset splits together with the geospatial contrastive loss applied directly to pairs of BEV feature grids.

If this is right

  • The semi-supervised model outperforms the supervised baseline on vectorized map perception tasks.
  • Qualitative segmentation of the BEV feature space improves under PCA visualization.
  • Labeling effort can be reduced by shifting supervision to a smaller single-traversal subset while adding unlabeled multi-traversal data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same overlap-analysis and contrastive mechanism could be applied to other BEV-based perception tasks such as object detection or lane segmentation.
  • Performance may continue to rise as the number of traversals per location increases beyond the minimum required by the current splits.
  • The method supplies a concrete route for scaling map-construction models to city-scale unlabeled fleets without proportional growth in annotation cost.

Load-bearing premise

The overlap-analysis procedure reliably yields contrastive pairs whose geospatial consistency directly improves the latent BEV representation for the downstream vectorized map task rather than merely fitting to dataset-specific traversal patterns.

What would settle it

Running the identical model and data splits with and without the contrastive term and finding no gain (or a loss) in vectorized map metrics together with no improvement in PCA-based segmentation of the BEV feature space.

Figures

Figures reproduced from arXiv: 2603.10688 by Alexander Blumberg, Christoph Stiller, Jan-Hendrik Pauls, Jonas Merkert.

Figure 1
Figure 1. Figure 1: Overlapping BEV feature grids visualized using PCA. In [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic overview of the semi-supervised learning pipeline. Data flows are shown for supervised (pink) and self-supervised (blue, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Single- and multi-traversals within Argoverse 2: The histogram on the left shows the general distribution of intersecting drive logs. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance scaling and relative gains across increasing [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative visualization of PCA of the supervised baseline (middle) and our semi-supervised approach (right) with ground-truth [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
read the original abstract

Autonomous vehicles rely on map information to understand the world around them. However, the creation and maintenance of offline high-definition (HD) maps remains costly. A more scalable alternative lies in online HD map construction, which only requires map annotations at training time. To further reduce the need for annotating vast training labels, self-supervised training provides an alternative. This work focuses on improving the latent birds-eye-view (BEV) feature grid representation within a vectorized online HD map construction model by enforcing geospatial consistency between overlapping BEV feature grids as part of a contrastive loss function. To ensure geospatial overlap for contrastive pairs, we introduce an approach to analyze the overlap between traversals within a given dataset and generate subsidiary dataset splits following adjustable multi-traversal requirements. We train the same model supervised using a reduced set of single-traversal labeled data and self-supervised on a broader unlabeled set of data following our multi-traversal requirements, effectively implementing a semi-supervised approach. Our approach outperforms the supervised baseline across the board, both quantitatively in terms of the downstream tasks vectorized map perception performance and qualitatively in terms of segmentation in the principal component analysis (PCA) visualization of the BEV feature space.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces MapGCLR, a semi-supervised approach for online vectorized HD map construction. It presents an overlap-analysis procedure to generate multi-traversal dataset splits and applies a contrastive loss to enforce geospatial consistency across overlapping BEV feature grids from different traversals. The same model architecture is trained supervised on a reduced single-traversal labeled subset and self-supervised on a broader unlabeled multi-traversal set; the abstract claims this yields quantitative gains on downstream vectorized map perception tasks and qualitative improvements visible in PCA visualizations of the BEV feature space relative to the supervised baseline.

Significance. If the central claim survives proper controls for data volume, the work would offer a concrete route to leverage unlabeled multi-traversal data for stronger BEV representations without extra map annotations, which is relevant for scalable online mapping in autonomous vehicles. The overlap-analysis procedure itself is a reusable engineering contribution for dataset construction. The paper receives credit for explicitly framing the protocol as semi-supervised and for evaluating on the downstream vectorized perception task rather than only on the contrastive objective.

major comments (1)
  1. [Abstract and methods (training protocol)] The experimental protocol trains the supervised baseline exclusively on a reduced single-traversal labeled subset while the proposed method consumes additional unlabeled multi-traversal data; no ablation is reported that equalizes total data volume across conditions or that replaces the overlap-derived positive/negative pairs with random or non-geospatial negatives. This comparison is described in the abstract and in the training-protocol paragraph of the methods. Because the reported outperformance cannot yet be isolated from the semi-supervised data regime, the attribution to geospatial consistency remains unsupported and is load-bearing for the central claim.
minor comments (1)
  1. [Abstract] The abstract states that the method 'outperforms the supervised baseline across the board' but supplies no numerical metrics, baseline names, or dataset statistics, which prevents immediate verification of the quantitative claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the need to better isolate the contribution of geospatial contrastive learning from the semi-supervised data regime. We address this point directly below and commit to adding the requested controls.

read point-by-point responses

  1. Referee: [Abstract and methods (training protocol)] The experimental protocol trains the supervised baseline exclusively on a reduced single-traversal labeled subset while the proposed method consumes additional unlabeled multi-traversal data; no ablation is reported that equalizes total data volume across conditions or that replaces the overlap-derived positive/negative pairs with random or non-geospatial negatives. This comparison is described in the abstract and in the training-protocol paragraph of the methods. Because the reported outperformance cannot yet be isolated from the semi-supervised data regime, the attribution to geospatial consistency remains unsupported and is load-bearing for the central claim.

    Authors: We agree that the current experimental design does not fully disentangle the effect of additional unlabeled data volume from the specific use of overlap-derived geospatial pairs. In the revised manuscript we will add two ablations: (1) a data-volume-matched supervised baseline trained on the union of the original labeled set plus pseudo-labels generated from the multi-traversal data, and (2) a contrastive variant that replaces the overlap-derived positive/negative pairs with randomly sampled pairs while keeping the same total data volume and loss weighting. These controls will allow us to quantify how much of the reported gain is attributable to geospatial consistency versus simply having access to more traversals. We will also update the abstract and methods to reflect these additional experiments. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain; method uses standard contrastive loss on externally generated pairs.

full rationale

The paper describes a semi-supervised pipeline that first computes geospatial overlaps from dataset traversals to form contrastive pairs, then applies a standard contrastive loss to BEV features before downstream supervised map construction. No equation or claim reduces the target performance metric to a fitted parameter by construction, nor does any load-bearing step rely on a self-citation whose content is itself unverified. The supervised baseline is trained on a reduced single-traversal subset while the contrastive branch consumes additional unlabeled multi-traversal data; this is an explicit experimental choice rather than a definitional identity. The derivation therefore remains self-contained against external data splits and loss functions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no equations, no modeling assumptions, and no experimental protocol, so the ledger remains empty.

pith-pipeline@v0.9.0 · 5754 in / 1092 out tokens · 26232 ms · 2026-05-25T06:42:40.274867+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 2 internal anchors

  1. [1]

    Autoware on board: Enabling autonomous ve- hicles with embedded systems,

    S. Kato, S. Tokunaga, Y . Maruyama, S. Maeda, M. Hirabayashi, Y . Kitsukawa, et al., “Autoware on board: Enabling autonomous ve- hicles with embedded systems,” in2018 ACM/IEEE 9th International Conference on Cyber-Physical Systems (ICCPS), 2018, pp. 287–296

    work page 2018

  2. [2]

    Lanelet2: A high-definition map framework for the future of automated driving,

    F. Poggenhans, J.-H. Pauls, J. Janosovits, S. Orf, M. Naumann, F. Kuhnt, et al., “Lanelet2: A high-definition map framework for the future of automated driving,” in2018 21st International Conference on Intelligent Transportation Systems (ITSC), Maui, HI, USA: IEEE Press, 2018, pp. 1672–1679,ISBN: 978-1-7281-0321-1

    work page 2018

  3. [3]

    Maptr: Structured modeling and learning for online vectorized hd map construction,

    B. Liao, S. Chen, X. Wang, T. Cheng, Q. Zhang, W. Liu, et al., “Maptr: Structured modeling and learning for online vectorized hd map construction,” inInternational Conference on Learning Representations, 2023

    work page 2023

  4. [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, et al., “Maptrv2: An end-to-end framework for online vectorized hd map construction,”International Journal of Computer Vision, pp. 1–23, 2024

    work page 2024

  5. [5]

    MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping,

    J. Chen, Y . Wu, J. Tan, H. Ma, and Y . Furukawa, “MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping,” inComputer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds., Cham: Springer Nature Switzerland, 2025, pp. 90–107,ISBN: 978-3-031- 72658-3

    work page 2024

  6. [6]

    M3tr: A generalist model for real-world hd map completion,

    F. Immel, R. Fehler, F. Bieder, J.-H. Pauls, and C. Stiller, “M3tr: A generalist model for real-world hd map completion,”IEEE Robotics and Automation Letters, vol. 10, no. 12, pp. 12 541–12 548, 2025

    work page 2025

  7. [7]

    Sdtagnet: Leveraging text-annotated navigation maps for online hd map construction,

    F. Immel, J.-H. Pauls, R. Fehler, F. Bieder, J. Merkert, and C. Stiller, “Sdtagnet: Leveraging text-annotated navigation maps for online hd map construction,” 2025

    work page 2025

  8. [8]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” inPro- ceedings of the 37th International Conference on Machine Learning, ser. ICML’20, vol. 119, JMLR.org, Jul. 2020, pp. 1597–1607

    work page 2020

  9. [9]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, et al., “Emerging properties in self-supervised vision transformers,” inProceedings of the International Conference on Computer Vision (ICCV), 2021

    work page 2021

  10. [10]

    Proposal learning for semi-supervised object detection,

    P. Tang, C. Ramaiah, Y . Wang, R. Xu, and C. Xiong, “Proposal learning for semi-supervised object detection,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Jan. 2021, pp. 2291–2301

    work page 2021

  11. [11]

    End-to-end semi-supervised object detection with soft teacher,

    M. Xu, Z. Zhang, H. Hu, J. Wang, L. Wang, F. Wei, et al., “End-to-end semi-supervised object detection with soft teacher,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Oct. 2021, pp. 3060–3069

    work page 2021

  12. [12]

    Label matching semi-supervised object detection,

    B. Chen, W. Chen, S. Yang, Y . Xuan, J. Song, D. Xie, et al., “Label matching semi-supervised object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2022, pp. 14 381–14 390

    work page 2022

  13. [13]

    Semi-supervised learning for visual bird’s eye view semantic segmentation,

    J. Zhu, L. Liu, Y . Tang, F. Wen, W. Li, and Y . Liu, “Semi-supervised learning for visual bird’s eye view semantic segmentation,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 9079–9085

    work page 2024

  14. [14]

    Occfeat: Self-supervised occupancy feature pre- diction for pretraining bev segmentation networks,

    S. Sirko-Galouchenko, A. Boulch, S. Gidaris, A. Bursuc, A. V obecky, P. Pérez, et al., “Occfeat: Self-supervised occupancy feature pre- diction for pretraining bev segmentation networks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, Jun. 2024, pp. 4493–4503

    work page 2024

  15. [15]

    Bevcon: Advancing bird’s eye view perception with contrastive learning,

    Z. Leng, J. Yang, Z. Ren, and B. Zhou, “Bevcon: Advancing bird’s eye view perception with contrastive learning,”IEEE Robotics and Automation Letters, vol. 10, no. 4, pp. 3158–3165, 2025

    work page 2025

  16. [16]

    Pseudomap- trainer: Learning online mapping without hd maps,

    C. Löwens, T. Funke, J. Xie, and A. P. Condurache, “Pseudomap- trainer: Learning online mapping without hd maps,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Oct. 2025, pp. 5263–5272

    work page 2025

  17. [17]

    Exploring semi- supervised learning for online mapping,

    A. Lilja, E. Wallin, J. Fu, and L. Hammarstrand, “Exploring semi- supervised learning for online mapping,”2025 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 2468–2478, 2024

    work page 2025

  18. [18]

    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,” in2022 International Conference on Robotics and Automation (ICRA), Philadelphia, PA, USA: IEEE Press, 2022, pp. 4628–4634

    work page 2022

  19. [19]

    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,” inInternational conference on machine learning, PMLR, 2023

    work page 2023

  20. [20]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in Computer Vision – ECCV 2020: 16th European Conference, Glas- gow, UK, August 23–28, 2020, Proceedings, Part I, Glasgow, United Kingdom: Springer-Verlag, 2020, pp. 213–229,ISBN: 978-3-030- 58451-1

    work page 2020

  21. [21]

    Leveraging enhanced queries of point sets for vectorized map construction,

    Z. Liu, X. Zhang, G. Liu, J. Zhao, and N. Xu, “Leveraging enhanced queries of point sets for vectorized map construction,” inEuropean Conference on Computer Vision, 2024

    work page 2024

  22. [22]

    Damap: Distance-aware mapnet for high quality hd map construction,

    J. Dong, C. Li, Y . Lin, J. Fu, S. Zhou, and N. Zheng, “Damap: Distance-aware mapnet for high quality hd map construction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025

    work page 2025

  23. [23]

    Streammapnet: Streaming mapping network for vectorized online hd map con- struction,

    T. Yuan, Y . Liu, Y . Wang, Y . Wang, and H. Zhao, “Streammapnet: Streaming mapping network for vectorized online hd map con- struction,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Jan. 2024, pp. 7356–7365

    work page 2024

  24. [24]

    Augmenting lane perception and topology understanding with standard definition navigation maps,

    K. Z. Luo, X. Weng, Y . Wang, S. Wu, J. Li, K. Q. Weinberger, et al., “Augmenting lane perception and topology understanding with standard definition navigation maps,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 4029– 4035

    work page 2024

  25. [25]

    Enhanc- ing vectorized map perception with historical rasterized maps,

    X. Zhang, G. Liu, Z. Liu, N. Xu, Y . Liu, and J. Zhao, “Enhanc- ing vectorized map perception with historical rasterized maps,” in European Conference on Computer Vision, 2024

    work page 2024

  26. [26]

    Y . Du, S. Yang, L. Wang, Z. Hou, C. Cai, Z. Tan, et al.,Rtmap: Real-time recursive mapping with change detection and localization,

    work page

  27. [27]

    arXiv:2507.00980 [cs.CV]

    work page arXiv

  28. [28]

    A. v. d. Oord, Y . Li, and O. Vinyals,Representation Learning with Contrastive Predictive Coding, arXiv:1807.03748 [cs], Jan. 2019

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  29. [29]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    work page 2016

  30. [30]

    Argoverse 2: Next generation datasets for self-driving per- ception and forecasting,

    B. Wilson, W. Qi, T. Agarwal, J. Lambert, J. Singh, S. Khandelwal, et al., “Argoverse 2: Next generation datasets for self-driving per- ception and forecasting,” inProceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021

    work page 2021

  31. [31]

    DINOv3

    O. Siméoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, et al.,DINOv3, 2025. arXiv:2508.10104 [cs.CV]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Impact of localization errors on label quality for online hd map construction,

    A. Blumberg, J. Merkert, R. Fehler, F. Immel, F. Bieder, J.-H. Pauls, et al., “Impact of localization errors on label quality for online hd map construction,” in2025 IEEE Intelligent V ehicles Symposium (IV), 2025, pp. 1833–1840

    work page 2025

  33. [33]

    co / datasets / nvidia / PhysicalAI - Autonomous - Vehicles, Access requires acceptance of the NVIDIA Autonomous Vehicle Dataset License Agreement, 2025

    NVIDIA Corporation,Physicalai-autonomous-vehicles dataset, https : / / huggingface . co / datasets / nvidia / PhysicalAI - Autonomous - Vehicles, Access requires acceptance of the NVIDIA Autonomous Vehicle Dataset License Agreement, 2025

    work page 2025