Pith. sign in

REVIEW 6 major objections 7 minor 50 references

The paper claims that fusing an explicit bird's-eye road-structure map with implicit geometric consistency among centerlines improves topological mapping accuracy on OpenLane-V2: +2.0% in OLS over the TopoLogic baseline and +5.7% mAP over L

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-01 07:53 UTC pith:PVWXEPAG

load-bearing objection Solid empirical paper on topological mapping with geometric priors, but the largest claimed gain is probably prior leakage and the attribution needs tighter ablations. the 6 major comments →

arxiv 2607.21281 v1 pith:PVWXEPAG submitted 2026-07-23 cs.CV cs.ROeess.IV

HGeo-TopoMap: Boosting Topological Mapping with Hierarchical Geometric Priors

Siyu Li , Kunyu Peng , Di Wen , Beiping Hou , Zhiyong Li , Kailun Yang This is my paper
classification cs.CV cs.ROeess.IV
keywords topological mappinglane topologygeometric priorsbird's-eye viewinverse perspective mappingcontrastive learningautonomous drivingOpenLane-V2
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper targets a specific failure in autonomous driving perception: road centerlines have no explicit visual markings, so detector models struggle to locate them and infer their connections. The authors propose to supply two complementary priors: an explicit road-structure map, obtained by lifting standard perspective segmentation into bird's-eye view, and implicit geometric regularities of centerlines themselves, such as straight-versus-curved shape and parallel spacing. They build two modules around these priors and show on the OpenLane-V2 benchmark that the combined system improves overall topology accuracy and lane-segment detection over strong baselines, and degrades more gracefully when camera views are missing. The paper also demonstrates the same explicit prior transfers to vectorized map construction, with a large reported gain in mAP. For a sympathetic reader, the contribution is a principled way to inject geometry that is cheap, model-agnostic, and testable.

Core claim

The central claim is that hierarchical geometric priors can compensate for the missing visual cues of centerlines in topological mapping. Concretely, on the OpenLane-V2 centerline benchmark the paper reports a +2.0% gain in overall topology score (OLS) over the TopoLogic baseline, with centerline detection up +1.6% and topology reasoning up +1.2%. On the lane-segment benchmark, it reports 34.0% mAP, a +5.7 gain over its re-implemented LaneSegNet baseline. The paper further reports that under missing-view robustness tests the method consistently beats baselines, and that the explicit prior alone lifts vectorized mapping by +14.5 mAP on the StreamMapNet baseline.

What carries the argument

The load-bearing mechanism is the pairing of two modules. Geometric Adaptive Learning (GAL) encodes a BEV road-structure map—computed by inverse perspective mapping from off-the-shelf perspective segmentation—into discrete semantic and positional tokens, then uses a distance-based prior-mask attention to let centerline queries attend only to nearby informative road-structure cells. Geometric Consistency Learning (GCL) first changes instance regression to a two-stage center-point-to-instance-point scheme with geometry-aware cross-attention, then applies an InfoNCE contrastive loss that pulls together centerline features sharing the same geometric group (straight parallel, straight perpendicul

Load-bearing premise

The paper's gains rest on the assumption that the BEV road-structure map produced by off-the-shelf segmentation plus inverse perspective mapping is reliable and non-redundant in the target scenes; if that map already encodes the classes being predicted, the reported improvements could come from leakage rather than from the geometric prior itself.

What would settle it

Corrupt the road-structure prior in a controlled way—randomly shuffle category labels, drop the divider and pedestrian classes, or jitter the map spatially—and rerun the same pipeline. If the OLS and mAP gains persist under corruption, the prior is not the cause; if they collapse, the method is only as good as the off-the-shelf segmentation and the flat-ground IPM assumption.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Centerline detection and topology reasoning on OpenLane-V2 improve by +1.6% DET_l and +1.2% TOP_ll respectively when both prior modules are active.
  • Lane-segment perception improves to 34.0% mAP, and inter-lane topology improves by +1.8%, relative to the LaneSegNet baseline.
  • Under missing-view camera conditions, the method maintains higher centerline detection and overall OLS than the baseline, with the largest gain (+7.5% DET_l) when the rear-left view is absent.
  • The explicit road-structure prior can be inserted into a different decoder (vectorized map construction), yielding a +14.5 mAP gain over StreamMapNet.
  • Ablation shows the two modules are complementary: GCL alone adds +0.9 DET_l / +0.8 OLS, and GAL adds a further +0.7 DET_l / +1.2 OLS.

Where Pith is reading between the lines

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

  • The +14.5 mAP gain on the vectorized-mapping task deserves caution: the road-structure prior already contains divider and pedestrian classes that StreamMapNet predicts, so part of the gain may be label leakage rather than learned geometric guidance; a clean test would remove those classes from the prior or corrupt them.
  • The method's sensitivity to prior-map quality is untested; since the authors themselves note in the conclusion that single-frame road-structure maps are information-deficient in complex scenes, evaluating with a weaker segmentation model or with synthetic spatial noise would reveal how much of the gain is genuine.
  • The geometric-group contrastive loss could generalize beyond centerlines to lane boundaries, crosswalks, or intersections, and temporal aggregation of road-structure maps—flagged by the authors as future work—is a natural extension that could stabilize the prior under occlusion.
  • Because the prior comes from a pre-trained segmentation model, the method inherits that model's domain bias; retraining or adapting the prior encoder on target-domain data is a testable extension.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 7 minor

Summary. The paper proposes HGeo-TopoMap, a method for topological mapping in BEV that incorporates hierarchical geometric priors. It consists of two modules: a Geometric Adaptive Learning (GAL) module, which processes a BEV road-structure map obtained by feeding off-the-shelf Mask2Former segmentations through inverse perspective mapping, and a Geometric Consistency Learning (GCL) module, which imposes geometric regularities on centerline instances via a center-point regression head and a contrastive loss over straight/curved groups. The method is evaluated on OpenLane-V2 for the centerline benchmark (OLS 45.5 vs. TopoLogic* 43.5), the lane-segment benchmark (mAP 34.0 vs. LaneSegNet* 28.3), a vectorized-mapping task (mAP 77.9 vs. StreamMapNet 63.4), and under missing-view robustness conditions. The central claim is that the proposed hierarchical prior guidance consistently improves topological mapping accuracy and robustness.

Significance. If the reported gains are attributable to the proposed mechanism, the paper makes a useful practical contribution: it shows how mature perspective semantic segmentation, combined with implicit geometric consistency, can compensate for the absence of explicit centerline visual cues. The paper's strengths are its broad evaluation across centerline, lane-segment, vectorized-map, and robustness settings, its ablation structure, and its commitment to releasing code and weights. However, the experimental attribution is currently under-supported because the BEV road-structure prior contains the very classes being predicted in the lane-segment and vectorized-map tasks, and because several ablation and hyperparameter choices are internally inconsistent. These issues do not invalidate the architecture, but they prevent the paper, in its present form, from establishing that the gains come from the claimed geometric-prior mechanism rather than from a strong external segmentation map.

major comments (6)
  1. [§IV-C, Table VII; §III-C] The prior-map leakage concern is central and unresolved. In the lane-segment benchmark (Table II), the BEV road-structure prior contains the pedestrian class, and AP_ped increases by +6.9 over the re-implemented baseline (35.7 vs. 28.8). In the vectorized-mapping experiment (Table VII), the prior map contains exactly the classes predicted (pedestrians, dividers, boundaries), and the method reports a +14.5 mAP gain, with AP_boun actually dropping 6.2 points. This pattern is consistent with the decoder reading target-class predictions from the external Mask2Former/IPM map rather than learning geometric priors. The paper provides no control: it does not ablate the semantic content of the prior map, does not compare with a trivial fusion baseline (e.g., concatenating the BEV map as an extra input channel), and does not quantify the prior map's precision/recall on the target domain. Please ad
  2. [§IV-B, §IV-D, Table IV] The final GCL loss weight is not the one selected by the ablation. Implementation details (§IV-B) set β=0.1, but Table IV (labelled “Weight of Geometric Consistency Learning Loss”) reports OLS 44.6 for β=0.01 and OLS 44.3 for β=0.1, i.e., 0.01 is the OLS-optimal value. The text justifies 0.1 by “more balanced performance”, but the table is not run on the full model (it appears to contain only the GCL module), and no sweep on the full model is reported. State explicitly whether Table IV uses the full HGeo-TopoMap or a GCL-only model, and report the β sweep on the final configuration.
  3. [§IV-D, Table V vs. Table II] The final lane-segment results do not match the ablation's recommended configuration. Table V shows that for the lane-segment baseline the best fusion position is layer 3, with mAP 33.7, AP_ls 31.7, AP_ped 35.8, TOP_lsls 22.4. The final reported Ours row in Table II is mAP 34.0, AP_ls 32.3, AP_ped 35.7, TOP_lsls 22.6, which coincides with no row in Table V. Please report the exact fusion layer and all hyperparameters used for the final model, and ensure Table V is consistent with the configuration whose numbers appear in Table II.
  4. [Table III] The ablation table does not include a GAL-only row. The text states that “both GCL and GAL contribute positively” and reports gains of +0.9% and +0.7%, but Table III only contains the baseline, GCL-only, and GCL+GAL rows. The GAL contribution is therefore measured only as an incremental gain over GCL, not in isolation. Since GAL is the module that introduces the external road-structure prior, a GAL-only row is essential to separate the effect of the explicit prior from the effect of GCL. Add this row to Table III.
  5. [§III-D] The geometric grouping used by the contrastive loss is not specified precisely. The text defines G={G_l1, G_l2, G_c} based on “geometric curvature” and “slopes of vector points”, but does not give the curvature threshold, the slope clustering algorithm, the number of points used, or the value of the reliability threshold t_s. Because the positive/negative sample selection for L_gcl depends entirely on these groupings, the method is not reproducible without these details. Please provide the exact computation or a precise algorithmic reference.
  6. [Tables I–VII; §IV-F] No error bars or repeated-seed statistics are reported. Several central improvements are small in absolute terms (e.g., +2.0 OLS over TopoLogic* in Table I, +0.8 OLS for GCL in Table III), and the robustness claims in Fig. 5 are presented without variance or a description of how missing-view inputs are generated. Given the small margins and the leakage concern above, please report standard deviations across at least three seeds for the main benchmarks, and state the exact missing-view protocol (which camera masks are applied, whether the model is retrained, and the scale of the reported 7.5% improvement).
minor comments (7)
  1. [§IV-D] The heading “Weight of Geometric Consistency Learning Loss” and the text refer to the weight factor α, but Eq. (16) uses α for the center-point loss and β for L_gcl. This is a typo; it should be β.
  2. [Table I] The Topo2D row is misaligned: “29.150.622.3 26.2” should be separated into DET_l=29.1, DET_t=50.6, TOP_ll=22.3, TOP_lt=26.2. Please fix the formatting.
  3. [Table VII] Column headings use “APped”, “APdiv”, “APboun” while the text uses AP_ped, AP_div, AP_boun. Make the notation uniform and add spaces.
  4. [§IV-F, Fig. 5] The missing-view robustness evaluation lacks a precise description of the protocol. State whether missing views are zeroed out or replaced, whether the model is trained with random missing views or tested only, and include numerical values or a table for the reported improvements.
  5. [§III-D] The sentence “straight centerlines can be broadly divided into two major groups” is vague. Clarify whether the grouping is based on the sign of the slope, the angle modulo π, or a clustering algorithm, and how the “slopes of vector points” are aggregated over a polyline.
  6. [References] Several references are to 2025/2026 proceedings or preprints. Please verify publication statuses and add arXiv identifiers where appropriate.
  7. [Fig. 2] The architecture diagram is dense; the query labels Q_t, Q_l and the contrastive-head path are hard to read. Increase font size or separate the GCL sub-block for clarity.

Circularity Check

0 steps flagged

No significant circularity: the paper reports empirical benchmark gains from an architecture with externally generated priors; no prediction reduces to its input by construction.

full rationale

The paper's central claims are empirical improvements on OpenLane-V2, not derivations from assumptions that contain the conclusions. The BEV road-structure prior is generated externally via Mask2Former and IPM in Sec. III-A, so it is not a fitted parameter and is not defined in terms of the benchmark outputs. The geometric consistency learning in Sec. III-D uses standard contrastive learning on the model's own matched queries; this is a training objective rather than a prediction that reduces to its input. The only self-citation, [12] for IPM, is non-load-bearing because IPM is a standard geometric transformation. The vectorized-map result in Table VII does show that the prior map contains the target classes, and the paper explicitly acknowledges this: 'given that the map contains divider and pedestrian classes that correspond to vectorized maps, integrating this prior yields a large boost of +14.5% in map accuracy.' This is a potential evaluation/attribution concern (possible leakage or trivial fusion), but it is not a circular derivation: the prior is not constructed from the benchmark labels and is not a renamed version of the method's own outputs. No step meets the evidence bar of Eq. X = Eq. Y by construction or fitted-parameter-renamed-as-prediction, so the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The method relies on several hand-tuned hyperparameters and assumptions about the validity of a pre-trained segmentation model and the geometric structure of roads. No new physical or conceptual entity is introduced.

free parameters (5)
  • k (prior-masked attention scaling factor) = 0.1
    Set by hand in Sec. IV-B; controls the softness of the attention mask.
  • alpha (center point loss weight) = 1
    Chosen without ablation; balances L_c in the total loss.
  • beta (geometric contrastive loss weight) = 0.1
    Chosen after ablation but the ablation suggests 0.01 gives better OLS; final 0.1 is justified by a 'balanced performance' statement rather than a clear rule.
  • instance reliability threshold t_s
    Used for selecting positive/negative samples in contrastive learning; value not specified, said to follow prior works.
  • Geometric grouping thresholds (curvature, slopes)
    The paper does not provide concrete values or algorithms for classifying centerlines as straight/curved and for splitting into slope groups; these are implicit free parameters.
axioms (4)
  • domain assumption IPM flat-ground assumption (Z=0)
    Sec. III-A2 assumes a flat road surface to project perspective maps into BEV; in real scenes with elevation changes this introduces distortion, which the paper acknowledges but does not quantify.
  • domain assumption Mask2Former pre-trained on Mapillary Vistas provides reliable perspective road maps
    The entire GAL module depends on this off-the-shelf segmentation being accurate enough; no uncertainty or failure analysis is provided.
  • domain assumption Centerlines can be meaningfully categorized into straight and curved, and straight centerlines are parallel/perpendicular
    Used to define the contrastive groups in Sec. III-D; if this geometric categorization is not discriminative in real data, the GCL loss may not help.
  • standard math The loss components from TopoLogic (L_det, L_l, L_topo) are valid for the new architecture
    The paper reuses baseline losses without re-deriving them; this is a standard practice but still an unstated assumption about their effectiveness.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of HGeo-TopoMap: Boosting Topological Mapping with Hierarchical Geometric Priors." pith.science (2026). https://pith.science/paper/PVWXEPAG

@misc{pith2026260721281,
  author       = {Pith},
  title        = {Pith review of: HGeo-TopoMap: Boosting Topological Mapping with Hierarchical Geometric Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PVWXEPAG}},
  note         = {Machine review of arXiv:2607.21281}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Topological maps are key outputs of autonomous driving perception systems, delivering essential road information for path planning. They identify instances such as centerlines and traffic signs, along with their connectivity relationships. Due to the lack of explicit markings for centerlines in real-world environments, the detection of centerline instances remains a significant challenge. To tackle this problem, we propose HGeo-TopoMap, which leverages an explicit prior map and implicit spatial relations to hierarchically boost topological mapping. First, a geometric adaptive learning module is designed for the road structure map obtained via inverse perspective mapping. This module discretely encodes semantic and spatial features from the map, followed by a prior-mask attention mechanism that selectively focuses on informative regions. Then, a geometric consistency learning module is devised, which leverages the geometric properties and spatial relationships of centerlines. Built on the geometry-aware decoder, it enforces spatial consistency by aligning features of centerline instances with identical geometric orientations. The proposed method is evaluated on the OpenLane-V2 dataset across the centerline, lane segment, and robustness benchmarks. Beyond substantial improvements in topological mapping accuracy, the proposed method offers the benefit of enhanced robustness, consistently outperforming baselines under both standard and challenging conditions. The source code and model weights will be made publicly available at https://github.com/lynn-yu/HGeo-TopoMap.

Figures

Figures reproduced from arXiv: 2607.21281 by Beiping Hou, Di Wen, Kailun Yang, Kunyu Peng, Siyu Li, Zhiyong Li.

Figure 1
Figure 1. Figure 1: Motivation of HGeo-TopoMap. To address the lack [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed HGeo-TopoMap architecture. The proposed geometric adaptive learning module embeds [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A comparison of instance point regression learn [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on two baselines of topologic mapping, TopoLogic [ [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparative evaluation of topological map quality [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

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

50 extracted references · 3 linked inside Pith

  1. [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,” inProc. ICRA, 2022, pp. 4628–4634

    2022

  2. [2]

    MapTRv2: An end-to-end framework for online vector- ized HD map construction,

    B. Liaoet al., “MapTRv2: An end-to-end framework for online vector- ized HD map construction,”International Journal of Computer Vision, vol. 133, no. 3, pp. 1352–1374, 2025

    2025

  3. [3]

    A concise survey on lane topology reasoning for HD mapping,

    Y . Yaoet al., “A concise survey on lane topology reasoning for HD mapping,” inProc. IV, 2025, pp. 468–475

    2025

  4. [4]

    HiVT: Hierarchical vector transformer for multi-agent motion prediction,

    Z. Zhou, L. Ye, J. Wang, K. Wu, and K. Lu, “HiVT: Hierarchical vector transformer for multi-agent motion prediction,” inProc. CVPR, 2022, pp. 8813–8823

    2022

  5. [5]

    DenseTNT: End-to-end trajectory predic- tion from dense goal sets,

    J. Gu, C. Sun, and H. Zhao, “DenseTNT: End-to-end trajectory predic- tion from dense goal sets,” inProc. ICCV, 2021, pp. 15 283–15 292

    2021

  6. [6]

    Graph-based topology reasoning for driving scenes,

    T. Liet al., “Graph-based topology reasoning for driving scenes,” Science China Information Sciences, vol. 69, no. 5, p. 152103, 2026

    2026

  7. [7]

    TopoLogic: An interpretable pipeline for lane topology reasoning on driving scenes,

    Y . Fuet al., “TopoLogic: An interpretable pipeline for lane topology reasoning on driving scenes,” inProc. NeurIPS, 2024, pp. 61 658– 61 676

    2024

  8. [8]

    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,” inProc. ECCV, 2020, pp. 213–229

    2020

  9. [9]

    LaneSegNet: Map learning with lane segment perception for autonomous driving,

    T. Liet al., “LaneSegNet: Map learning with lane segment perception for autonomous driving,” inProc. ICLR, 2024

    2024

  10. [10]

    Masked-attention mask transformer for universal image segmentation,

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-attention mask transformer for universal image segmentation,” inProc. CVPR, 2022, pp. 1280–1289

    2022

  11. [11]

    SegFormer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “SegFormer: Simple and efficient design for semantic segmentation with transformers,” inProc. NeurIPS, 2021, pp. 12 077–12 090

    2021

  12. [12]

    GenMapping: Unleashing the potential of inverse perspective mapping for robust online HD map construction,

    S. Li, K. Yang, H. Shi, S. Wang, Y . Yao, and Z. Li, “GenMapping: Unleashing the potential of inverse perspective mapping for robust online HD map construction,”IEEE Open Journal of Intelligent Transportation Systems, vol. 7, pp. 1492–1506, 2026

    2026

  13. [13]

    OpenLane-V2: A topology reasoning benchmark for unified 3D HD mapping,

    H. Wanget al., “OpenLane-V2: A topology reasoning benchmark for unified 3D HD mapping,” inProc. NeurIPS, 2023, pp. 18 873–18 884

    2023

  14. [14]

    Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3D,

    J. Philion and S. Fidler, “Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3D,” inProc. ECCV, 2020, pp. 194–210

    2020

  15. [15]

    BEVFormer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers,

    Z. Liet al., “BEVFormer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers,” inProc. ECCV, 2022, pp. 1–18

    2022

  16. [16]

    Bird’s-eye-view panoptic segmentation using monocular frontal view images,

    N. Gosala and A. Valada, “Bird’s-eye-view panoptic segmentation using monocular frontal view images,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 1968–1975, 2022

    1968

  17. [17]

    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 Proc. ICLR, 2021

    2021

  18. [18]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” inProc. CVPR, 2016, pp. 779–788

    2016

  19. [19]

    MapTR: Structured modeling and learning for online vectorized HD map construction,

    B. Liaoet al., “MapTR: Structured modeling and learning for online vectorized HD map construction,” inProc. ICLR, 2023

    2023

  20. [20]

    StreamMapNet: Streaming mapping network for vectorized online HD map construc- tion,

    T. Yuan, Y . Liu, Y . Wang, Y . Wang, and H. Zhao, “StreamMapNet: Streaming mapping network for vectorized online HD map construc- tion,” inProc. WACV, 2024, pp. 7341–7350

    2024

  21. [21]

    PriorDrive: Enhancing online HD mapping with unified vector priors,

    S. Zenget al., “PriorDrive: Enhancing online HD mapping with unified vector priors,” inProc. AAAI, 2026, pp. 12 313–12 321

    2026

  22. [22]

    Topo2Seq: Enhanced topology reasoning via topology sequence learning,

    Y . Yanget al., “Topo2Seq: Enhanced topology reasoning via topology sequence learning,” inProc. AAAI, 2025, pp. 9318–9326

    2025

  23. [23]

    TopoBDA: Towards bezier deformable attention for road topology understanding,

    M. E. Kalfaoglu, H. I. Ozturk, O. Kilinc, and A. Temizel, “TopoBDA: Towards bezier deformable attention for road topology understanding,” Neurocomputing, p. 132360, 2025

    2025

  24. [24]

    TopoPoint: Enhance topology reasoning via endpoint detection in autonomous driving,

    Y . Fu, X. Liu, T. Li, Y . Ma, Y . Zhang, and F. Dai, “TopoPoint: Enhance topology reasoning via endpoint detection in autonomous driving,” in Proc. NeurIPS, 2025, pp. 118 229–118 250

    2025

  25. [25]

    TopoHR: Hierarchical centerline representation for cyclic topology reasoning in driving scenes with point-to-instance relations,

    Y . Bai, Z. Chen, B. Song, E. Cheng, and H. Ling, “TopoHR: Hierarchical centerline representation for cyclic topology reasoning in driving scenes with point-to-instance relations,” inProc. CVPR, 2026, pp. 18 161– 18 170

    2026

  26. [26]

    Reusing attention for one-stage lane topology understand- ing,

    Y . Liet al., “Reusing attention for one-stage lane topology understand- ing,” inProc. IROS, 2025, pp. 16 977–16 984

    2025

  27. [27]

    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,” inProc. ICLR, 2024, pp. 45 604–45 615

    2024

  28. [28]

    RATopo: Improving lane topology reasoning via redundancy assignment,

    H. Li, S. Huang, L. Xu, Y . Gao, B. Mu, and S. Liu, “RATopo: Improving lane topology reasoning via redundancy assignment,” inProc. MM, 2025, pp. 777–786

    2025

  29. [29]

    T2SG: Traffic topology scene graph for topology reasoning in autonomous driving,

    C. Lv, M. Qi, L. Liu, and H. Ma, “T2SG: Traffic topology scene graph for topology reasoning in autonomous driving,” inProc. CVPR, 2025, pp. 17 197–17 206

    2025

  30. [30]

    RelTopo: Multi-level relational modeling for driving scene topology reasoning,

    Y . Luoet al., “RelTopo: Multi-level relational modeling for driving scene topology reasoning,”arXiv preprint arXiv:2506.13553, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    TopoMaskV3: 3D mask head with dense offset and height predictions for road topology understanding,

    M. E. Kalfaoglu, H. I. Öztürk, O. Kilinc, and A. Temizel, “TopoMaskV3: 3D mask head with dense offset and height predictions for road topology understanding,” inProc. CVPRW, 2026, pp. 675–684

    2026

  32. [32]

    Chameleon: Fast-slow neuro-symbolic lane topology extraction,

    Z. Zhanget al., “Chameleon: Fast-slow neuro-symbolic lane topology extraction,” inProc. ICRA, 2025, pp. 3752–3758

    2025

  33. [33]

    Persistent autoregressive mapping with traffic rules for autonomous driving,

    S. Lianget al., “Persistent autoregressive mapping with traffic rules for autonomous driving,” inProc. AAAI, 2026, pp. 6862–6870

    2026

  34. [34]

    Exploring real world map change generalization of prior-informed HD map prediction models,

    S. M. Batemanet al., “Exploring real world map change generalization of prior-informed HD map prediction models,” inProc. CVPRW, 2024, pp. 4568–4578

    2024

  35. [35]

    GlobalMapNet: An online framework for vectorized global HD map construction,

    A. Shi, Y . Cai, X. Chen, J. Pu, Z. Fu, and H. Lu, “GlobalMapNet: An online framework for vectorized global HD map construction,”arXiv preprint arXiv:2409.10063, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    P-MapNet: Far-seeing map generator enhanced by both SDMap and HDMap priors,

    Z. Jianget al., “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

  37. [37]

    DiffMap: Enhancing map segmentation with map prior using diffusion model,

    P. Jiaet al., “DiffMap: Enhancing map segmentation with map prior using diffusion model,”IEEE Robotics and Automation Letters, vol. 9, no. 11, pp. 9836–9843, 2024

    2024

  38. [38]

    SMART: Advancing scalable map priors for driving topology reasoning,

    J. Yeet al., “SMART: Advancing scalable map priors for driving topology reasoning,” inProc. ICRA, 2025, pp. 3298–3304

    2025

  39. [39]

    SEPT: Standard-definition map enhanced scene perception and topology reasoning for autonomous driving,

    M. Peiet al., “SEPT: Standard-definition map enhanced scene perception and topology reasoning for autonomous driving,”IEEE Robotics and Automation Letters, vol. 10, no. 7, pp. 7126–7133, 2025

    2025

  40. [40]

    Coherent online road topology estimation and reasoning with standard-definition maps,

    K. S. Pham, C. Witte, J. Behley, J. Betz, and C. Stachniss, “Coherent online road topology estimation and reasoning with standard-definition maps,” inProc. IROS, 2025, pp. 9886–9893

    2025

  41. [41]

    Enhancing lane segment perception and topology reasoning with crowdsourcing trajectory priors,

    P. Jiaet al., “Enhancing lane segment perception and topology reasoning with crowdsourcing trajectory priors,”IEEE Robotics and Automation Letters, vol. 10, no. 6, pp. 5417–5424, 2025

    2025

  42. [42]

    Generating synthetic deviation maps for prior-enhanced vectorized HD map construction,

    H. Xu, Y . Xiao, W. Li, and Y . Hu, “Generating synthetic deviation maps for prior-enhanced vectorized HD map construction,” inProc. IV, 2025, pp. 419–426

    2025

  43. [43]

    AMap: Distilling future priors for ahead-aware online HD map construction,

    R. Liet al., “AMap: Distilling future priors for ahead-aware online HD map construction,” inProc. CVPR, 2026, pp. 24 906–24 917

    2026

  44. [44]

    D2HDMap: Non-visible driveline map prior for online vectorized HD map prediction,

    S. Shon, C. Tsuchiya, D. Bhanderi, D. Ilstrup, H. Cheng, and C. Ostafew, “D2HDMap: Non-visible driveline map prior for online vectorized HD map prediction,” inProc. IV, 2026

    2026

  45. [45]

    Aerial- FusionMapNet: Online HD map construction with aerial-onboard BEV fusion,

    D. Lengerer, M. Pechinger, K. Bogenberger, and C. Markgraf, “Aerial- FusionMapNet: Online HD map construction with aerial-onboard BEV fusion,” inProc. ITSC, 2026

    2026

  46. [46]

    MGMap: Mask- guided learning for online vectorized HD map construction,

    X. Liu, S. Wang, W. Li, R. Yang, J. Chen, and J. Zhu, “MGMap: Mask- guided learning for online vectorized HD map construction,” inProc. CVPR, 2024, pp. 14 812–14 821

    2024

  47. [47]

    Improving hierarchical representations of vectorized HD maps with perspective clues,

    C. Zhang, Q. Song, F. Li, J. Li, and R. Huang, “Improving hierarchical representations of vectorized HD maps with perspective clues,”IEEE Robotics and Automation Letters, vol. 10, no. 12, pp. 12 533–12 540, 2025

    2025

  48. [48]

    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, 2024

    Pith/arXiv arXiv 2024

  49. [49]

    Uncertainty-aware panoptic segmentation,

    K. Sirohi, S. Marvi, D. Büscher, and W. Burgard, “Uncertainty-aware panoptic segmentation,”IEEE Robotics and Automation Letters, vol. 8, no. 5, pp. 2629–2636, 2023

    2023

  50. [50]

    The mapillary vistas dataset for semantic understanding of street scenes,

    G. Neuhold, T. Ollmann, S. R. Bulò, and P. Kontschieder, “The mapillary vistas dataset for semantic understanding of street scenes,” inProc. ICCV, 2017, pp. 5000–5009

    2017

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.