Pith. sign in

REVIEW 4 major objections 5 minor 62 references

Coherent Online Road Topology Estimation and Reasoning with Standard-Definition Maps

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Score jointly predicts lane segments, road boundaries, traffic elements, and their topology in one network, using SD map priors and temporal fusion, reaching OLUS 54.9 on OpenLane-V2 subset A.

desk verdict A solid, well-ablated systems contribution to SD-map-assisted online lane/topology perception; the main weakness is the missing robustness evaluation under imperfect SD maps, not the architecture itself. read the letter →

arxiv 2507.01397 v2 pith:UNQWPTYK submitted 2025-07-02 cs.CV cs.LG

Khanh Son Pham , Christian Witte , Jens Behley , Johannes Betz , Cyrill Stachniss This is my paper
classification cs.CVcs.LG
keywords SDmappriorslanesegmentdetectionroadtopologyestimationtrafficelementassociationtemporalfusionquerydenoisingBEVperceptionOpenLane-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

This paper proposes Score, a single network that constructs an online road model from multi-view camera images plus a standard-definition (SD) map, predicting lane segments, road boundaries, and 2D traffic elements and reasoning about lane-to-lane and traffic-element-to-lane associations. The authors argue that prior map information supplies strong initial hypotheses for where lanes are, which helps especially when lanes are occluded, and that denoising and temporal fusion make training stable and outputs consistent across frames. On the OpenLane-V2 subset A validation split, Score reaches an OLUS of 54.9, up from 47.5 for Topologic and 36.7 for LaneSegNet, and it also improves on a geographically disjoint split. If correct, this shows that SD maps, which are widely available and cheap, can substantially reduce the need for expensive HD maps in autonomous driving perception.

What carries the argument

The load-bearing mechanism is the SD-map-enhanced lane segment query. For each of $n_{sd}=50$ queries, a reference point is sampled from the middle of an SD map edge, with additional points distributed by edge length, and the query embedding is formed as $\hat{p}_{sd,i}=\mathrm{MLP}(q_{pos,i})+p_{sd,i}$, combining a learned offset with the map prior. These queries initialize the lane decoder, are masked so that erroneous SD map information cannot leak to other queries, and are remapped by the decoder against BEV features. Around this, the paper adds lane denoising queries from noisy ground-truth polylines (DN-DETR style), a lane attention that distributes reference points along predicted boundaries, a ConvGRU temporal fusion of BEV features, and a post-processing topology score that blends learned similarity with a distance-based term $A^D_{lsls}(i,j)=2/(1+\exp(d_{i,j}/\alpha))$.

What would settle it

Artificially corrupt the SD map input on the OpenLane-V2 validation set, for example by deleting 20% of OSM polylines at random and shifting another 20% laterally by 1–2 meters, then re-run the trained Score model. If the OLUS score stays near 54.9, the SD prior is not load-bearing; if it drops substantially, the paper's claim that map priors drive the improvement is confirmed.

Watch

Extended reading notes

Core claim

Score's central claim is that coherent online HD-map-style perception can be formulated as a single query-based transformer pipeline in which lane segment queries are initialized from SD map polylines, refined against bird's-eye-view features, and used to predict lane geometry, lane boundaries, road boundaries, and traffic elements, with topology heads reading out adjacency matrices. The key result is that this coherent formulation beats existing centerline- or segment-based methods on every OpenLane-V2 subset A task: DETls 44.0 vs 37.0 for TopoSD, DETa 43.4, DETte 61.4, TOPll 40.0, TOPlt 39.1, giving OLUS 54.9 on validation and 55.8 on test. The paper attributes the gains to three mechanisms: SD-enhanced queries that sample reference points from map edges, lane denoising groups that stabilize bipartite matching, and recurrent temporal fusion of BEV features with ego-motion compensation.

Load-bearing premise

The load-bearing premise is that the SD map polylines used to sample reference points are topologically correct and roughly aligned with the lanes the camera sees; if a map lane is missing or misaligned, the query starts from a wrong location and the decoder may never recover.

Editorial extensions

If this is right

  • SD map priors let the model predict occluded lane geometry, since map-derived queries provide hypotheses even when visual evidence is missing.
  • Using OSM maps instead of OpenLane-V2's own SD annotations improves results, indicating that the quality of the prior map matters.
  • Temporal fusion adds a substantial gain (TOPll from 36.2 to 40.0), suggesting that online streaming is important for topology consistency.
  • A single coherent model outperforms separate or staged approaches on all OpenLane-V2 subset A tasks, reducing the motivation for task-specific ensembles.
  • The method generalizes to geographically disjoint splits better than LaneSegNet, so the gains are not merely memorized locations.

Reading between the lines

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

  • If SD-map topology is wrong or misaligned, performance should degrade sharply; the paper's own masking of SD queries suggests the authors anticipate this, and a stress test deleting map segments would quantify how much of the 54.9 is prior versus vision.
  • The approach could be ported to nuScenes subset B once lane segment annotations exist, since SD priors are available there; current evaluation is only subset A.
  • The dependence on OSM quality implies that deployment in areas with sparse OSM data would need a fallback that detects when the prior is unreliable, for example by monitoring the learned offset magnitude.
  • The OLUS gap between validation (54.9) and test (55.8) with different DET distributions suggests that per-task weighting could shift design priorities; on the test split TOPlt (42.2) outweighs TOPll (34.5).
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 / 5 minor

Summary. The paper proposes Score, a multi-task network that jointly predicts lane segments, road boundaries, 2D traffic elements, lane-lane topology, and traffic-element-to-lane associations on the OpenLane-V2 subset A benchmark. The main technical elements are SD-map-enhanced lane queries whose reference points are initialized from sampled SD-map polylines (Eq. 2), SMERF-style cross-attention of SD-map features into BEV features, lane denoising queries, one-to-many matching, dataset resampling, and ConvGRU-based temporal fusion. The topology head combines a learned pairwise score with a distance-based post-processing term (Eqs. 5-7). Experiments report OLUS 54.9 versus 47.5 for Topologic and 36.7 for LaneSegNet on the OpenLane-V2 validation split, and an improved DETls on a geographically disjoint split (Table III).

Significance. If the results are reproducible, Score demonstrates that a single coherent architecture can leverage SD-map priors to outperform prior lane-segment and centerline methods by a large margin on a standard benchmark. The disjoint-split evaluation is a genuine strength because it addresses the location-leakage issue identified in earlier mapping datasets. The ablation study in Table II attributes concrete gains to each component, and the empirical gains are consistent across the main metric components. The main limitations are external validity: the benchmark provides perfectly aligned SD maps in the ego frame, the comparison set excludes the two Mapless Driving track methods that use the same task formulation, and all headline numbers are single-run without error bars. These limitations are addressable in revision rather than being fatal to the approach.

major comments (4)
  1. [Section III-A.1, Eq. (2)] The SD-map-enhanced queries are initialized at points sampled from SD-map polylines, and SMERF fuses SD-map geometry into the BEV features; in OpenLane-V2 the SD map is provided in the ego-vehicle frame with perfect localization. The paper does not evaluate the effect of realistic SD-map perturbations, such as 1-2 m localization shifts, missing edges, or stale/mislabeled segments. Since the self-attention mask only prevents deficient SD information from leaking to other queries and does not correct misplaced initial reference points, the practical claim of "reduces the dependence on HD maps" is not yet established. Please add a robustness experiment with perturbed SD maps, or explicitly restrict the claim to conditions with well-aligned SD maps.
  2. [Section IV-C, Table I] The "state-of-the-art" claim is made against a comparison set that excludes LGMap and MapVision, the two SD-map-based Mapless Driving track entries. The stated justification (larger backbones, test-time augmentation, ensembles) is legitimate, but it means the reported "large margin" does not cover the strongest published methods on the exact same task. Please report those methods' scores in a separate row or table, even if marked as not directly comparable, and ideally add a same-backbone ablation so readers can separate architectural contribution from additional compute.
  3. [Section III-D, Eqs. (6)-(7)] The distance-based topology term AD and its fusion weight β are tuned on the validation set, with only the selected values α=2.5 and β=0.8 reported. Table II shows that this post-processing adds +1.4 TOPll before temporal fusion, so its contribution to the headline topology gain is non-negligible. Please include a sensitivity table over α and β, or retune these hyperparameters on the training split and report validation performance, so the topology claim is not dependent on validation-set tuning.
  4. [Tables I and III] All reported metrics are single-run and no variance estimates are provided, and the disjoint-split comparison uses different training schedules (LaneSegNet at 12/24 epochs versus Score at 15/30 temporal-fusion epochs). The margins in Table I are large, so the single-run issue may be less severe for the validation comparison, but the current format does not allow the reader to judge whether the Table III gains are meaningful. Please report at least two seeds or standard deviations for the main comparison, or clearly state the run-to-run variability observed.
minor comments (5)
  1. [Section III-A.2 and IV-B] The lane denoising scale λ_dn is defined but its numerical value is never reported in the experimental setup; please provide it.
  2. [Section III-E.3] The wheel-angle resampling threshold of 7° is given, but the resampling factor, the resulting training-set size, and the final class balance are not reported; please quantify these choices.
  3. [Section III-E.4 and Table VI] The warm-up phase is described only by reference to prior work; please specify the number of warm-up epochs, when the ConvGRU recurrence is enabled during training, and whether validation is run in single-frame or recurrent mode.
  4. [Table I] The validation block contains two LaneSegNet rows, one with DETls=32.3 and no OLUS and one marked with an asterisk with DETls=27.8 and OLUS=36.7; clarify the source and exact configuration of each row so that the claimed 36% improvement is anchored to a well-defined baseline.
  5. [Section IV-A, Eq. (11)] The functions f(TOPll) and f(TOPlt) in the OLUS formula are not defined; please provide the exact benchmark transformation so the aggregation can be reproduced.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Score's SD-map prior is an external input; lane/topology outputs are supervised by ground truth and independent learned components.

full rationale

Score's derivation chain is self-contained: the SD map is an external input, not a restatement of the output. Lane queries are initialized from sampled SD-map points (Eq. 2) but are refined by a learned decoder supervised with Hungarian matching and explicit vector, segmentation, classification, and topology losses (Eq. 8); topology is not read off the SD map. The distance-based post-processing (Eqs. 6-7) is a deterministic function of the model's own predicted endpoints, applied with tuned constants alpha=2.5 and beta=0.8, and is combined with a learned adjacency score A^S trained by focal loss, so it does not reduce the topology output to its input by construction. The 2D traffic element detector (YOLOv9) is a separately trained external component and is transparently reported. Self-citations (e.g., references [39] and [47]) appear only in background enumerations of related work and are not load-bearing for any central claim. The geographically disjoint split (Table III) and ablations against external baselines provide independent evidence. The benchmark's perfectly aligned SD maps are a robustness/domain-gap limitation for real deployment, not a circularity in the derivation.

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

The central claim rests on standard domain assumptions about dataset quality, SD map alignment, and BEV feature sufficiency. The only hand-tuned numbers that directly enter the output formulas are alpha and beta in the topology post-processing; training loss weights and query counts are architectural hyperparameters that are not specific to the scientific claim.

free parameters (4)
  • alpha (topology distance scale) = 2.5
    Set by empirical analysis in Sec. III-D; controls how quickly the distance-based topology score decays in Eq. (6).
  • beta (topology fusion weight) = 0.8
    Set by empirical analysis in Sec. III-D; weights the distance-based topology term against the learned score in Eq. (7).
  • lambda_dn (denoising rescale) = not stated (lambda_dn in (0,1))
    Hyperparameter in Eq. (3) that rescales noisy reference points; no specific value is reported in the paper.
  • wheel angle resampling threshold = 7 degrees
    Threshold in Sec. III-E3 for upsampling turning samples from roughly 4% of the dataset; chosen by dataset statistics.
assumptions (4)
  • domain assumption OpenLane-V2 subset A annotations provide accurate and complete ground truth for lane segments, road boundaries, and topology.
    All training and evaluation metrics depend on these labels; annotation noise or incompleteness would bias DET and TOP scores.
  • domain assumption SD map (OSM or OpenLane-V2) polylines are topologically correct and sufficiently aligned with the visual scene to serve as lane hypotheses.
    Reference points for SD-enhanced queries are sampled from these polylines (Sec. III-A1); missing or misaligned map data would initialize queries at incorrect locations.
  • domain assumption BEVFormer's bird's-eye-view features provide sufficient geometric information for lane, boundary, and topology decoding.
    Lane and boundary decoders cross-attend to F_BEV from BEVFormer (Sec. III-A, III-B); degraded BEV features would cap downstream performance.
  • domain assumption LaneSegNet's lane segment formulation, Hungarian matching, and loss weighting are valid for the joint task.
    Score reuses LaneSegNet's decoder, matching costs, and loss terms (Sec. III-A, III-F), so any flaw in that base would propagate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Coherent Online Road Topology Estimation and Reasoning with Standard-Definition Maps." pith.science (2026). https://pith.science/paper/UNQWPTYK

@misc{pith2026250701397,
  author       = {Pith},
  title        = {Pith review of: Coherent Online Road Topology Estimation and Reasoning with Standard-Definition Maps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UNQWPTYK}},
  note         = {Machine review of arXiv:2507.01397}
}
read the original abstract

Most autonomous cars rely on the availability of high-definition (HD) maps. Current research aims to address this constraint by directly predicting HD map elements from onboard sensors and reasoning about the relationships between the predicted map and traffic elements. Despite recent advancements, the coherent online construction of HD maps remains a challenging endeavor, as it necessitates modeling the high complexity of road topologies in a unified and consistent manner. To address this challenge, we propose a coherent approach to predict lane segments and their corresponding topology, as well as road boundaries, all by leveraging prior map information represented by commonly available standard-definition (SD) maps. We propose a network architecture, which leverages hybrid lane segment encodings comprising prior information and denoising techniques to enhance training stability and performance. Furthermore, we facilitate past frames for temporal consistency. Our experimental evaluation demonstrates that our approach outperforms previous methods by a large margin, highlighting the benefits of our modeling scheme.

Figures

Figures reproduced from arXiv: 2507.01397 by the authors.

Figure 1
Figure 1. Given multi-view images and a standard-definition (SD) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of our approach. First, multi-view image features are transformed into a BEV representation. Utilizing SD map [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results. Our method demonstrates the capacity to incorporate a SD map prior and past information in an effective manner, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 53 canonical work pages

  1. [1]

    Caesar, V

    H. Caesar, V . Bankiti, A.H. Lang, S. V ora, V .E. Liong, Q. Xu, A. Kr- ishnan, Y . Pan, G. Baldan, and O. Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2020

    work page 2020

  2. [2]

    Y .B. Can, A. Liniger, D.P. Paudel, and L. Van Gool. Structured bird’s- eye-view traffic scene understanding from onboard images. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021

    work page 2021

  3. [3]

    Carion, F

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko. End-to-end object detection with transformers. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2020

    work page 2020

  4. [4]

    J. Chen, Y . Wu, J. Tan, H. Ma, and Y . Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2024

    work page 2024

  5. [5]

    S. Chen, T. Cheng, X. Wang, W. Meng, Q. Zhang, and W. Liu. Efficient and robust 2d-to-bev representation learning via geometry- guided kernel transformer. arXiv preprint, arXiv:2206.04584, 2022

    arXiv 2022

  6. [6]

    S. Choi, J. Kim, H. Shin, and J.W. Choi. Mask2Map: Vectorized HD Map Construction Using Bird’s Eye View Segmentation Masks. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2024

    work page 2024

  7. [7]

    Chung, C

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling. In Proc. of the Neural Information Processing Systems Workshops , 2014

    work page 2014

  8. [8]

    W. Ding, L. Qiao, X. Qiu, and C. Zhang. Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2023

    work page 2023

Show all 62 references
  1. [9]

    Y . Fu, W. Liao, X. Liu, Y . Ma, F. Dai, and Y . Zhang. TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes. In Proc. of the Conf. on Neural Information Processing Systems (NeurIPS), 2024

    2024

  2. [10]

    C. Han, J. Yang, J. Sun, Z. Ge, R. Dong, H. Zhou, W. Mao, Y . Peng, and X. Zhang. Exploring recurrent long-term temporal fusion for multi-view 3d perception. IEEE Robotics and Automation Letters (RA- L), 9(7):6544–6551, 2024

    2024

  3. [11]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning for Image Recognition. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) , 2016

    2016

  4. [12]

    Huang and G

    J. Huang and G. Huang. Bevdet4d: Exploit temporal cues in multi- camera 3d object detection. arXiv preprint, arXiv:2203.17054, 2022

    2022 arXiv

  5. [13]

    Huang, G

    J. Huang, G. Huang, Z. Zhu, Y . Ye, and D. Du. Bevdet: High- performance multi-camera 3d object detection in bird-eye-view. arXiv preprint, arXiv:2112.11790, 2021

    2021 arXiv

  6. [14]

    D. Jia, Y . Yuan, H. He, X. Wu, H. Yu, W. Lin, L. Sun, C. Zhang, and H. Hu. Detrs with hybrid matching. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2023

    2023

  7. [15]

    J. Jiao. Machine learning assisted high-definition map creation. In Proc. of the IEEE Annual Computer Software and Applications Conf. (COMPSAC), 2018

    2018

  8. [16]

    Kingma and J

    D. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In Proc. of the Intl. Conf. on Learning Representations (ICLR) , 2015

    2015

  9. [17]

    F. Li, H. Zhang, S. Liu, J. Guo, L.M. Ni, and L. Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  10. [18]

    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, 2024

    2024 arXiv

  11. [19]

    Q. Li, Y . Wang, Y . Wang, and H. Zhao. Hdmapnet: An online hd map construction and evaluation framework. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA) , 2022

    2022

  12. [20]

    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, 2023

    2023 arXiv

  13. [21]

    T. Li, P. Jia, B. Wang, L. Chen, K. Jiang, J. Yan, and H. Li. Lanesegnet: Map learning with lane segment perception for autonomous driving. In Proc. of the Intl. Conf. on Learning Representations (ICLR) , 2024

    2024

  14. [22]

    Y . Li, Z. Ge, G. Yu, J. Yang, Z. Wang, Y . Shi, J. Sun, and Z. Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proc. of the Conf. on Advancements of Artificial Intelligence (AAAI), 2023

    2023

  15. [23]

    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 Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2022

    2022

  16. [24]

    B. Liao, S. Chen, B. Jiang, T. Cheng, Q. Zhang, W. Liu, C. Huang, and X. Wang. Lane graph as path: Continuity-preserving path- wise modeling for online lane graph construction. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2024

    2024

  17. [25]

    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 Proc. of the Intl. Conf. on Learning Representations (ICLR), 2023

    2023

  18. [26]

    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. Intl. Journal of Computer Vision (IJCV), 133(3):1352–1374, 2024

    2024

  19. [27]

    Lilja, J

    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 Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  20. [28]

    T.Y . Lin, P. Doll ´ar, R. Girshick, K. He, B. Hariharan, and S. Be- longie. Feature pyramid networks for object detection. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  21. [29]

    T.Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar. Focal Loss for Dense Object Detection. In Proc. of the IEEE Intl. Conf. on Computer Vision (ICCV), 2017

    2017

  22. [30]

    X. Lin, T. Lin, Z. Pei, L. Huang, and Z. Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint, arXiv:2211.10581, 2022

    2022 arXiv

  23. [31]

    X. Lin, Z. Pei, T. Lin, L. Huang, and Z. Su. Sparse4d v3: Advancing end-to-end 3d detection and tracking. arXiv preprint , arXiv:2311.11722, 2023

    2023 arXiv

  24. [32]

    X. Liu, S. Wang, W. Li, R. Yang, J. Chen, and J. Zhu. Mgmap: Mask- guided learning for online vectorized hd map construction. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  25. [33]

    Y . Liu, T. Yuan, Y . Wang, Y . Wang, and H. Zhao. Vectormapnet: End-to-end vectorized hd map learning. In Proc. of the Intl. Conf. on Machine Learning (ICML) , 2023

    2023

  26. [34]

    Y . Liu, T. Wang, X. Zhang, and J. Sun. Petr: Position embedding transformation for multi-view 3d object detection. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2022

    2022

  27. [35]

    Y . Liu, J. Yan, F. Jia, S. Li, A. Gao, T. Wang, and X. Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV) , 2023

    2023

  28. [36]

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D.L. Rus, and S. Han. Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2023

    2023

  29. [37]

    K.Z. Luo, X. Weng, Y . Wang, S. Wu, J. Li, K.Q. Weinberger, Y . Wang, and M. Pavone. Augmenting lane perception and topology understanding with standard definition navigation maps. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA) , 2024

    2024

  30. [38]

    Z. Ma, S. Liang, Y . Wen, W. Lu, and G. Wan. RoadPainter: Points Are Ideal Navigators for Topology transformER. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2024

    2024

  31. [39]

    Opra, B.L

    B. Opra, B.L. Dem, J.M. Walls, D. Lukarski, and C. Stachniss. Leveraging GNSS and Onboard Visual Data from Consumer Vehicles for Robust Road Network Estimation. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , 2024

    2024

  32. [40]

    Philion and S

    J. Philion and S. Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2020

    2020

  33. [41]

    L. Qiao, W. Ding, X. Qiu, and C. Zhang. End-to-end vectorized hd-map construction with piecewise bezier curve. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  34. [42]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, L. Kaiser, and I. Polosukhin. Attention Is All You Need. In Proc. of the Conf. Neural Information Processing Systems (NIPS) , 2017

    2017

  35. [43]

    Wang and H.Y .M

    C.Y . Wang and H.Y .M. Liao. YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2024

    2024

  36. [44]

    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 Proc. of the Conf. on Neural Information Processing Systems (Neu...

    2023

  37. [45]

    S. Wang, F. Jia, Y . Liu, Y . Zhao, Z. Chen, T. Wang, C. Zhang, X. Zhang, and F. Zhao. Stream query denoising for vectorized hd map construction. In Proc. of the Europ. Conf. on Computer Vision (ECCV), 2024

    2024

  38. [46]

    Wilson, W

    B. Wilson, W. Qi, T. Agarwal, J. Lambert, J. Singh, S. Khandelwal, B. Pan, R. Kumar, A. Hartnett, J.K. Pontes, D. Ramanan, P. Carr, and J. Hays. Argoverse 2: Next Generation Datasets for Self- driving Perception and Forecasting. In Proc. of the Conf. on Neural Information Proc...

    2021

  39. [47]

    Witte, J

    C. Witte, J. Behley, C. Stachniss, and M. Raaijmakers. Epipolar Atten- tion Field Transformers for Bird’s Eye View Semantic Segmentation. In Proc. of the IEEE Winter Conf. on Applications of Computer Vision (WACV), 2025

    2025

  40. [48]

    D. Wu, J. Chang, F. Jia, Y . Liu, T. Wang, and J. Shen. TopoMLP: An Simple yet Strong Pipeline for Driving Topology Reasoning. In Proc. of the Intl. Conf. on Learning Representations (ICLR) , 2024

    2024

  41. [49]

    K. Wu, S. Nian, C. Shen, C. Yang, and Z. Li. LGmap: Local-to- Global Mapping Network for Online Long-Range Vectorized HD Map Construction. arXiv preprint, arXiv:2406.13988, 2024

    2024 arXiv

  42. [50]

    Z. Xia, X. Pan, S. Song, L.E. Li, and G. Huang. Vision transformer with deformable attention. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2022

    2022

  43. [51]

    Z. Xu, Y . Liu, Y . Sun, M. Liu, and L. Wang. Centerlinedet: Centerline graph detection for road lanes with vehicle-mounted sensors by transformer for hd map generation. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA) , 2023

    2023

  44. [52]

    Xu, K.Y .K

    Z. Xu, K.Y .K. Wong, and H. Zhao. InsMapper: Exploring Inner- instance Information for Vectorized HD Mapping. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2024

    2024

  45. [53]

    S. Yang, M. Jiang, Z. Fan, X. Xie, X. Tan, Y . Li, E. Ding, L. Wang, and J. Wang. TopoSD: Topology-Enhanced Lane Segment Perception with SDMap Prior. arXiv preprint, arXiv:2411.14751, 2024

    2024 arXiv

  46. [54]

    Z. Yang, J. Liu, Z. Wu, P. Wu, and X. Liu. Video Event Restoration Based on Keyframes for Video Anomaly Detection. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  47. [55]

    Z. Yang, M. Liu, J. Xie, Y . Zhang, C. Shen, W. Shao, J. Jiao, T. Xing, R. Hu, and P. Xu. MapVision: CVPR 2024 Autonomous Grand Challenge Mapless Driving Tech Report. arXiv preprint , arXiv:2406.10125, 2024

    2024 arXiv

  48. [56]

    J. Yu, Z. Zhang, S. Xia, and J. Sang. Scalablemap: Scalable map learning for online long-range vectorized hd map construction. arXiv preprint, arXiv:2310.13378, 2023

    2023 arXiv

  49. [57]

    T. Yuan, Y . Liu, Y . Wang, Y . Wang, and H. Zhao. Streammapnet: Streaming mapping network for vectorized online hd map construc- tion. In Proc. of the IEEE Winter Conf. on Applications of Computer Vision (WACV), 2024

    2024

  50. [58]

    Zhang, D

    H. Zhang, D. Paz, y. Guo, A. Das, X. Huang, K. Haug, H. Chris- tensen, and L. Ren. Enhancing Online Road Network Perception and Reasoning with Standard Definition Maps. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS) , 2024

    2024

  51. [59]

    Zhang, Y

    Z. Zhang, Y . Zhang, X. Ding, F. Jin, and X. Yue. Online vectorized hd map construction using geometry. In Proc. of the Europ. Conf. on Computer Vision (ECCV) , 2024

    2024

  52. [60]

    Zhou and P

    B. Zhou and P. Kr ¨ahenb¨uhl. Cross-view transformers for real-time map-view semantic segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2022

    2022

  53. [61]

    H. Zhou, Z. Ge, Z. Li, and X. Zhang. Matrixvt: Efficient multi-camera to bev transformation for 3d perception. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV) , 2023

    2023

  54. [62]

    Y . Zhou, H. Zhang, J. Yu, Y . Yang, S. Jung, S.I. Park, and B. Yoo. HIMap: HybrId Representation Learning for End-to-end Vectorized HD Map Construction. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.