pith. sign in

arxiv: 2606.31814 · v1 · pith:TVPJPHEDnew · submitted 2026-06-30 · 💻 cs.CV

Generative Lane Topology Reasoning via Autoregressive Model with Geometry Prior

Jiahui Fu , Zehao Huang , Han Li , Naiyan Wang , Si Liu This is my paper

Pith reviewed 2026-07-01 06:07 UTC · model grok-4.3

classification 💻 cs.CV
keywords lane topology reasoningautoregressive modelgeometry priorlane graphOpenLane-V2generative frameworkperception adapterBEV features
0
0 comments X

The pith

An autoregressive transformer pre-trained on millions of map scenes learns lane graph geometry priors that transfer via a perception adapter to produce more consistent and complete lane topologies from sensor data.

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

The paper establishes a generative framework called TopoGPT that learns typical lane graph structures by serializing them into token sequences and training an autoregressive model to predict the next token conditioned on scene context. A large map dataset of 3.3 million scenes supplies the training data for this geometry prior. Existing detection-plus-association methods produce endpoint inconsistencies and missing segments because they handle lanes independently and cannot recover from occlusions. The pre-trained model is then adapted to real multi-view camera inputs by aligning bird's-eye-view features with the learned scene tokens, allowing the geometry prior to guide prediction. On the OpenLane-V2 benchmark this yields average gains of 6.4 points on lane-level metrics and 11.6 points on point-level metrics while producing geometrically consistent and structurally complete graphs.

Core claim

Pre-training an autoregressive lane sequence transformer on tokenized lane graphs from 3.3 million map scenes via scene-conditioned next-token prediction endows the model with a geometry prior over lane structures; a perception adapter then aligns BEV features extracted from multi-view images with the pre-trained scene condition, transferring the prior to produce lane graphs that are geometrically consistent at endpoints and structurally complete despite occlusions.

What carries the argument

Autoregressive lane sequence transformer pre-trained via scene-conditioned next-token prediction on tokenized lane graphs, plus a perception adapter that aligns BEV features to the pre-trained scene tokens.

If this is right

  • Lane graphs exhibit geometric consistency at connected endpoints rather than independent per-lane errors.
  • Graphs remain structurally complete even when individual lanes are occluded in the input images.
  • Average performance rises by 6.4 points on lane-level metrics and 11.6 points on point-level metrics on OpenLane-V2.
  • The same pre-training plus adaptation pipeline can be applied to other sensor modalities once their BEV features are aligned to the scene tokens.

Where Pith is reading between the lines

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

  • Similar autoregressive pre-training on large map corpora could supply priors for other structured prediction tasks such as traffic-sign topology or drivable-area connectivity.
  • The 3.3-million-scene map dataset may serve as a reusable resource for learning additional geometric or topological priors beyond lane graphs.
  • Because generation is sequential, inference latency may increase with the number of lanes; parallel decoding or caching strategies would be needed for real-time use.

Load-bearing premise

The geometry prior learned from map data transfers effectively to real sensor observations through the perception adapter without major loss of effectiveness.

What would settle it

Training the same architecture from scratch on OpenLane-V2 without the 3.3-million-scene map pre-training step and measuring whether lane-level and point-level scores drop to the level of prior detection-association methods.

Figures

Figures reproduced from arXiv: 2606.31814 by Han Li, Jiahui Fu, Naiyan Wang, Si Liu, Zehao Huang.

Figure 1
Figure 1. Figure 1: Samples of predicted lane graphs. Given predicted lane graphs, human intuition can readily distinguish realistic structures from unrealistic ones without visual input. Test your intuition. 1 Abstract. Lane topology reasoning aims to construct a lane graph from onboard sensor observations. Existing methods follow a detection and as￾sociation paradigm that treats each lane instance independently, leading to … view at source ↗
Figure 2
Figure 2. Figure 2: Given multi-view images: (a) Previous discriminative methods suffer from end￾points misalignment and missed lanes under occlusion. (b) Our generative model learns the geometry prior from large-scale map data and predicts contiguous and complete lane graphs conditioned on image-derived perception tokens. 1 Introduction Lane topology reasoning [43,56] constructs a vectorized lane graph from onboard sensor ob… view at source ↗
Figure 3
Figure 3. Figure 3: Framework of TopoGPT. Built on an autoregressive lane sequence trans￾former, we first perform geometric map-prior pre-training on large-scale map data to learn the geometry prior, where lane token sequences are generated conditioned on scene tokens \protect \textbf {X}_{\text {scene}} . We then conduct perception-aware alignment fine-tuning on sensor-map paired data, initializing from the pre-trained weigh… view at source ↗
Figure 4
Figure 4. Figure 4: Token sequence construction. We first build a unified lane graph from large-scale map data through a data processing pipeline. For each sampled scene, we (1) rasterize it and encode with a scene context encoder to obtain scene tokens, (2) convert each lane into a discrete group using a Bézier-based lane tokenizer and then sort all groups by a spatial lexicographical order to form the lane token sequence. b… view at source ↗
Figure 5
Figure 5. Figure 5: (a) During fine-tuning, \protect \mathbf {X}_{\text {bev}} is optimized under alignment and lane prediction objectives. (b) During inference, \protect \mathbf {X}_{\text {bev}} is directly used as condition for lane prediction. Building on this property, we propose a flow-matching-based adapter that evolves BEV features toward the pre-trained scene token distribution. We for￾mulate the procedure as a cross… view at source ↗
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison between TopoPoint and our TopoGPT. The first row shows multi-view inputs, and the second row denotes the lane graphs. Dark solid arrows indicate centerlines, while light dashed arrows depict topological relations between centerlines. ing, where the output features of the DiT model are directly used as \protect \mathbf {X}_{\text {bev}} and aligned with \protect \mathbf {X}_{\text {sc… view at source ↗
read the original abstract

Lane topology reasoning aims to construct a lane graph from onboard sensor observations. Existing methods follow a detection and association paradigm that treats each lane instance independently, leading to geometric inconsistency at connected endpoints and incomplete graphs due to visual occlusions. To address these issues, we propose TopoGPT, a generative framework that learns the geometry prior from typical lane graph structures through autoregressive sequence modeling. Specifically, we construct a large-scale map dataset comprising 3.3M scenes. For each lane graph, a lane tokenizer serializes it into discrete tokens, while a scene context encoder converts it into a rasterized image and extracts global features as scene tokens. We pre-train an autoregressive lane sequence transformer via scene-conditioned next-token prediction, endowing the model with the geometry prior over lane graph structures. Building upon this prior, a perception adapter aligns BEV features from multi-view images with the pre-trained scene condition, transferring the learned geometry prior to sensor-based lane graph prediction. On the OpenLane-V2 benchmark, TopoGPT outperforms existing methods by an average of +6.4 on lane-level and +11.6 on point-level metrics, and produces geometrically consistent and structurally complete lane graphs.

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

2 major / 2 minor

Summary. The paper proposes TopoGPT, a generative framework for lane topology reasoning from onboard sensors. It pre-trains an autoregressive transformer on 3.3M map scenes via scene-conditioned next-token prediction after tokenizing lane graphs and encoding rasterized scene context, thereby learning a geometry prior over lane structures. A perception adapter then aligns BEV features extracted from multi-view images with the pre-trained scene conditioning to transfer the prior, yielding lane graphs that are claimed to be geometrically consistent and structurally complete. On OpenLane-V2 the method reports average gains of +6.4 on lane-level metrics and +11.6 on point-level metrics over prior detection-and-association baselines.

Significance. If the transfer of the autoregressive prior succeeds without substantial degradation, the work offers a principled way to mitigate endpoint inconsistency and occlusion-induced incompleteness that plague independent-lane detectors. The scale of the map pre-training corpus (3.3M scenes) and the explicit separation of prior learning from perception are notable strengths that could influence future topology-reasoning pipelines in autonomous driving.

major comments (2)
  1. [perception adapter and experimental results] The central claim that the learned geometry prior transfers effectively through the perception adapter rests on the assumption that adapter-aligned BEV features lie in the same token space as the rasterized map scene tokens. No ablation that isolates the contribution of the pre-trained autoregressive transformer from the adapter architecture or from other modeling choices is described; without such a control it remains unclear whether the reported +6.4 / +11.6 gains derive from the generative prior or from the adapter itself.
  2. [perception adapter] The manuscript states that the adapter 'aligns BEV features … with the pre-trained scene condition,' yet provides no quantitative measure (e.g., token-space distance, reconstruction error on held-out map scenes, or distribution-shift statistics) demonstrating that the alignment preserves the next-token prediction behavior learned during pre-training.
minor comments (2)
  1. [introduction / method overview] The abstract and introduction would benefit from an explicit statement of the lane tokenizer vocabulary size and the rasterization resolution used for scene context encoding.
  2. [figures] Figure captions should clarify whether the visualized lane graphs are produced by the full TopoGPT pipeline or by an oracle adapter; this affects interpretation of geometric consistency.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the scale of our pre-training corpus and the separation of prior learning from perception. We address the two major comments below and will revise the manuscript to strengthen the evidence for the perception adapter.

read point-by-point responses

  1. Referee: The central claim that the learned geometry prior transfers effectively through the perception adapter rests on the assumption that adapter-aligned BEV features lie in the same token space as the rasterized map scene tokens. No ablation that isolates the contribution of the pre-trained autoregressive transformer from the adapter architecture or from other modeling choices is described; without such a control it remains unclear whether the reported +6.4 / +11.6 gains derive from the generative prior or from the adapter itself.

    Authors: We agree that an ablation isolating the pre-trained autoregressive transformer is required to substantiate the contribution of the geometry prior. In the revised manuscript we will add experiments that train the full pipeline from scratch (without map pre-training) and compare against the pre-trained TopoGPT, thereby quantifying how much of the reported gains is attributable to the transferred prior versus the adapter architecture alone. revision: yes

  2. Referee: The manuscript states that the adapter 'aligns BEV features … with the pre-trained scene condition,' yet provides no quantitative measure (e.g., token-space distance, reconstruction error on held-out map scenes, or distribution-shift statistics) demonstrating that the alignment preserves the next-token prediction behavior learned during pre-training.

    Authors: We acknowledge the absence of quantitative alignment diagnostics. In the revision we will report next-token prediction accuracy on held-out map scenes when the scene-conditioning tokens are replaced by adapter-aligned BEV features, together with a simple distribution-shift statistic (e.g., mean cosine distance in the token embedding space), to verify that the pre-trained next-token behavior is largely retained. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external map pre-training and benchmark evaluation

full rationale

The paper pre-trains an autoregressive transformer on a separately constructed 3.3M-scene map dataset using scene-conditioned next-token prediction to learn a geometry prior, then applies a perception adapter to transfer it to BEV features from multi-view images. Evaluation occurs on the external OpenLane-V2 benchmark with reported gains of +6.4 lane-level and +11.6 point-level. No self-definitional steps, fitted inputs called predictions, load-bearing self-citations, or uniqueness theorems appear in the derivation chain; the adapter and benchmark provide independent transfer and measurement steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract does not provide enough detail to identify any free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5747 in / 1102 out tokens · 35793 ms · 2026-07-01T06:07:40.631599+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

56 extracted references · 7 canonical work pages · 4 internal anchors

  1. [1]

    In: CVPR (2023)

    Büchner, M., Zürn, J., Todoran, I., Valada, A., Burgard, W.: Learning and aggre- gating lane graphs for urban automated driving. In: CVPR (2023)

    2023

  2. [2]

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

    Caesar, H., Kabzan, J., Tan, K.S., Fong, W.K., Wolff, E.M., Lang, A.H., Fletcher, L., Beijbom, O., Omari, S.: nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  3. [3]

    In: ICCV (2021) 16 J

    Can, Y.B., Liniger, A., Paudel, D.P., Van Gool, L.: Structured bird’s-eye-view traffic scene understanding from onboard images. In: ICCV (2021) 16 J. Fu et al

    2021

  4. [4]

    In: ICLR (2022)

    Chen, T., Saxena, S., Li, L., Fleet, D.J., Hinton, G.E.: Pix2seq: A language mod- eling framework for object detection. In: ICLR (2022)

    2022

  5. [5]

    In: ECCV (2024)

    Chitta, K., Dauner, D., Geiger, A.: SLEDGE: synthesizing driving environments with generative models and rule-based traffic. In: ECCV (2024)

    2024

  6. [6]

    In: ICCV (2021)

    Ettinger, S., Cheng, S., Caine, B., Liu, C., Zhao, H., Pradhan, S., Chai, Y., Sapp, B., Qi, C.R., Zhou, Y., Yang, Z., Chouard, A., Sun, P., Ngiam, J., Vasudevan, V., McCauley, A., Shlens, J., Anguelov, D.: Large scale interactive motion forecasting for autonomous driving : The waymo open motion dataset. In: ICCV (2021)

    2021

  7. [7]

    In: ICRA (2024)

    Fu, J., Gao, C., Wang, Z., Yang, L., Wang, X., Mu, B., Liu, S.: Eliminating cross- modal conflicts in bev space for lidar-camera 3d object detection. In: ICRA (2024)

    2024

  8. [8]

    In: CVPR (2025)

    Fu, J., Gong, Y., Wang, L., Zhang, S., Zhou, X., Liu, S.: Generative map priors for collaborative bev semantic segmentation. In: CVPR (2025)

    2025

  9. [9]

    In: NeurIPS (2024)

    Fu, Y., Liao, W., Liu, X., Xu, H., Ma, Y., Zhang, Y., Dai, F.: Topologic: An interpretable pipeline for lane topology reasoning on driving scenes. In: NeurIPS (2024)

    2024

  10. [10]

    In: NeurIPS (2025)

    Fu, Y., Liu, X., Li, T., Ma, Y., Zhang, Y., Dai, F.: Topopoint: Enhance topology reasoning via endpoint detection in autonomous driving. In: NeurIPS (2025)

    2025

  11. [11]

    In: ICRA (2024)

    Gao, W., Fu, J., Shen, Y., Jing, H., Chen, S., Zheng, N.: Complementing onboard sensors with satellite maps: A new perspective for hd map construction. In: ICRA (2024)

    2024

  12. [12]

    In: ECCV (2024)

    Han, Y., Yu, K., Li, Z.: Continuity preserving online centerline graph learning. In: ECCV (2024)

    2024

  13. [13]

    Harley, A.W., Fang, Z., Li, J., Ambrus, R., Fragkiadaki, K.: Simple-bev: What really matters for multi-sensor bev perception? In: ICRA (2023)

    2023

  14. [14]

    In: ICCV (2025)

    He, J., Yu, Q., Liu, Q., Chen, L.C.: Flowtok: Flowing seamlessly across text and image tokens. In: ICCV (2025)

    2025

  15. [15]

    In: CVPR (2016)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)

    2016

  16. [16]

    In: WACV (2022)

    He, S., Balakrishnan, H.: Lane-level street map extraction from aerial imagery. In: WACV (2022)

    2022

  17. [17]

    IEEE RAL (2025)

    Jia, P., Luo, Z., Wen, T., Yang, M., Jiang, K., Cui, L., Yang, D.: Enhancing lane segment perception and topology reasoning with crowdsourcing trajectory priors. IEEE RAL (2025)

    2025

  18. [18]

    IEEE RAL (2024)

    Jia, P., Wen, T., Luo, Z., Yang, M., Jiang, K., Liu, Z., Tang, X., Lei, Z., Cui, L., Zhang, B., et al.: Diffmap: Enhancing map segmentation with map prior using diffusion model. IEEE RAL (2024)

    2024

  19. [19]

    IEEE RAL (2024)

    Jiang, Z., Zhu, Z., Li, P., Gao, H., Yuan, T., Shi, Y., Zhao, H., Zhao, H.: P-mapnet: Far-seeing map generator enhanced by both sdmap and hdmap priors. IEEE RAL (2024)

    2024

  20. [20]

    In: ECCV (2024)

    Le, D.T., Shi, H., Cai, J., Rezatofighi, H.: Diffusion model for robust multi-sensor fusion in 3d object detection and bev segmentation. In: ECCV (2024)

    2024

  21. [21]

    IEEE TCSVT (2026)

    Li, H., Gao, Y., Liu, S., Wang, Y., Liu, B., Mu, B.: Unified modeling of lane and lane topology for driving scene reasoning. IEEE TCSVT (2026)

    2026

  22. [22]

    In: ACM MM (2025)

    Li,H.,Huang,S.,Xu,L.,Gao,Y.,Mu,B.,Liu,S.:Ratopo:Improvinglanetopology reasoning via redundancy assignment. In: ACM MM (2025)

    2025

  23. [23]

    arXiv preprint arXiv:2406.03105 (2024)

    Li, H., Huang, Z., Wang, Z., Rong, W., Wang, N., Liu, S.: Enhancing 3d lane detec- tion and topology reasoning with 2d lane priors. arXiv preprint arXiv:2406.03105 (2024)

    work page arXiv 2024

  24. [24]

    arXiv preprint arXiv:2304.05277 (2023) Generative Lane Topology Reasoning with Geometry Prior 17

    Li, T., Chen, L., Wang, H., Li, Y., Yang, J., Geng, X., Jiang, S., Wang, Y., Xu, H., Xu, C., et al.: Graph-based topology reasoning for driving scenes. arXiv preprint arXiv:2304.05277 (2023) Generative Lane Topology Reasoning with Geometry Prior 17

    work page arXiv 2023

  25. [25]

    In: ECCV (2024)

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

    2024

  26. [26]

    In: CVPR (2025)

    Liu, Q., Yin, X., Yuille, A.L., Brown, A., Singh, M.: Flowing from words to pixels: A noise-free framework for cross-modality evolution. In: CVPR (2025)

    2025

  27. [27]

    In: ICML (2023)

    Liu, Y., Yuan, T., Wang, Y., Wang, Y., Zhao, H.: Vectormapnet: End-to-end vec- torized hd map learning. In: ICML (2023)

    2023

  28. [28]

    IEEE TPAMI (2025)

    Lu, J., Nie, M., Zhang, B., Peng, R., Cai, X., Xu, H., Li, H., Wen, F., Zhang, W., Zhang, L.: Translating images to road network: A sequence-to-sequence perspec- tive. IEEE TPAMI (2025)

    2025

  29. [29]

    In: ICCV (2023)

    Lu, J., Peng, R., Cai, X., Xu, H., Li, H., Wen, F., Zhang, W., Zhang, L.: Translating images to road network: A non-autoregressive sequence-to-sequence approach. In: ICCV (2023)

    2023

  30. [30]

    In: ICRA (2024)

    Luo,K.Z.,Weng,X.,Wang,Y.,Wu,S.,Li,J.,Weinberger,K.Q.,Wang,Y.,Pavone, M.: Augmenting lane perception and topology understanding with standard defi- nition navigation maps. In: ICRA (2024)

    2024

  31. [31]

    In: CVPR (2025)

    Lv, C., Qi, M., Liu, L., Ma, H.: T2sg: Traffic topology scene graph for topology reasoning in autonomous driving. In: CVPR (2025)

    2025

  32. [32]

    In: ECCV (2024)

    Ma, Z., Liang, S., Wen, Y., Lu, W., Wan, G.: Roadpainter: Points are ideal navi- gators for topology transformer. In: ECCV (2024)

    2024

  33. [33]

    In: IROS (2025)

    Monninger, T., Zhang, Z., Mo, Z., Anwar, M.Z., Staab, S., Ding, S.: Mapdiffusion: Generative diffusion for vectorized online hd map construction and uncertainty estimation in autonomous driving. In: IROS (2025)

    2025

  34. [34]

    IEEE RAL (2022)

    Ort, T., Walls, J.M., Parkison, S.A., Gilitschenski, I., Rus, D.: Maplite 2.0: Online hd map inference using a prior sd map. IEEE RAL (2022)

    2022

  35. [35]

    In: ICCV (2023)

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023)

    2023

  36. [36]

    IEEE RAL (2025)

    Pei, M., Shan, J., Li, P., Shi, J., Huo, J., Gao, Y., Shen, S.: SEPT: standard- definition map enhanced scene perception and topology reasoning for autonomous driving. IEEE RAL (2025)

    2025

  37. [37]

    In: AAAI (2024)

    Peng, R., Cai, X., Xu, H., Lu, J., Wen, F., Zhang, W., Zhang, L.: Lanegraph2seq: Lane topology extraction with language model via vertex-edge encoding and con- nectivity enhancement. In: AAAI (2024)

    2024

  38. [38]

    In: IROS (2025)

    Pham, K.S., Witte, C., Behley, J., Betz, J., Stachniss, C.: Coherent online road topology estimation and reasoning with standard-definition maps. In: IROS (2025)

    2025

  39. [39]

    In: ACM MM (2024)

    Rong, F., Peng, W., Lan, M., Zhang, Q., Zhang, L.: Driving scene understanding with traffic scene-assisted topology graph transformer. In: ACM MM (2024)

    2024

  40. [40]

    In: CVPR (2025)

    Rowe, L., Girgis, R., Gosselin, A., Paull, L., Pal, C., Heide, F.: Scenario dreamer: Vectorized latent diffusion for generating driving simulation environments. In: CVPR (2025)

    2025

  41. [41]

    Autoregressive Model Beats Diffusion: Llama for Scalable Image Generation

    Sun, P., Jiang, Y., Chen, S., Zhang, S., Peng, B., Luo, P., Yuan, Z.: Autoregres- sive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  42. [42]

    In: NeurIPS (2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)

    2017

  43. [43]

    In: NeurIPS (2023)

    Wang, H., Li, T., Li, Y., Chen, L., Sima, C., Liu, Z., Wang, B., Jia, P., Wang, Y., Jiang, S., Wen, F., Xu, H., Luo, P., Yan, J., Zhang, W., Li, H.: Openlane-v2: A topology reasoning benchmark for unified 3d HD mapping. In: NeurIPS (2023)

    2023

  44. [44]

    In: NeurIPS (2024) 18 J

    Wang, L., Zhao, Y., Zhang, Z., Feng, J., Liu, S., Kang, B.: Image understanding makes for a good tokenizer for image generation. In: NeurIPS (2024) 18 J. Fu et al

    2024

  45. [45]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  46. [46]

    In: ICCV (2025)

    Wang, Z., Zhang, W., Zhang, W., Tan, X., Liu, H., Wang, Y., Li, G.: Lanediffusion: Improving centerline graph learning via prior injected BEV feature generation. In: ICCV (2025)

    2025

  47. [47]

    In: ICCV (2023)

    Wang, Z., Huang, Z., Fu, J., Wang, N., Liu, S.: Object as query: Lifting any 2d object detector to 3d detection. In: ICCV (2023)

    2023

  48. [48]

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

    Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  49. [49]

    In: ICLR (2024)

    Wu, D., Chang, J., Jia, F., Liu, Y., Wang, T., Shen, J.: Topomlp: A simple yet strong pipeline for driving topology reasoning. In: ICLR (2024)

    2024

  50. [50]

    In: ICCV (2025)

    Xie, M., Zeng, S., Chang, X., Liu, X., Pan, Z., Xu, M., Wei, X.: Seqgrowgraph: Learning lane topology as a chain of graph expansions. In: ICCV (2025)

    2025

  51. [51]

    In: ICRA (2023)

    Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Centerlinedet: Centerline graph de- tection for road lanes with vehicle-mounted sensors by transformer for hd map generation. In: ICRA (2023)

    2023

  52. [52]

    In: AAAI (2025)

    Yang, Y., Luo, Y., He, B., Li, E., Cao, Z., Zheng, C., Mei, S., Li, Z.: Topo2seq: Enhanced topology reasoning via topology sequence learning. In: AAAI (2025)

    2025

  53. [53]

    In: ICRA (2025)

    Ye, J., Paz, D., Zhang, H., Guo, Y., Huang, X., Christensen, H.I., Wang, Y., Ren, L.: Smart: Advancing scalable map priors for driving topology reasoning. In: ICRA (2025)

    2025

  54. [54]

    arXiv preprint arXiv:2409.05352 (2024)

    Zeng, S., Chang, X., Liu, X., Yuan, Y., Liang, S., Pan, Z., Xu, M., Wei, X.: Pri- ordrive: Enhancing online hd mapping with unified vector priors. arXiv preprint arXiv:2409.05352 (2024)

    work page arXiv 2024

  55. [55]

    In: ICCV (2023)

    Zhu, X., Zyrianov, V., Liu, Z., Wang, S.: Mapprior: Bird’s-eye view map layout estimation with generative models. In: ICCV (2023)

    2023

  56. [56]

    IEEE RAL (2021)

    Zürn, J., Vertens, J., Burgard, W.: Lane graph estimation for scene understanding in urban driving. IEEE RAL (2021)

    2021