Pith. sign in

REVIEW 3 major objections 7 minor 33 references

SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims lane topology can be learned as a chain of graph expansions, and that an autoregressive transformer trained on that chain reaches the best Landmark and Reachability F1 reported on nuScenes and Argoverse 2.

desk verdict A genuinely new lane-graph serialization with a plausible SOTA claim, but the evaluation protocol in Section 4.5 is ambiguous enough that the headline numbers should not be taken at face value. read the letter →

arxiv 2507.04822 v1 pith:NXAWF3FF submitted 2025-07-07 cs.CV

classification cs.CV
keywords lanetopologygraphconstructionautoregressivesequencegenerationexpansiontransformerdecoderBéziercurvesHDmapdepth-firstsearch
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

SeqGrowGraph claims that lane topology is best learned not by detecting centerlines and then reasoning about their connections, but by serializing the whole road graph into a chain of expansion steps and asking a transformer to predict that chain. At each step the model introduces one intersection node, declares which existing nodes it comes from and goes to, and appends quadratic Bézier control points for the connecting centerlines, so adjacency and geometry grow together. Trained on camera images through a bird's-eye-view encoder and a transformer decoder, the model autoregressively generates the complete graph. This matters because it gives HD map construction a unified representation that handles loops and bidirectional lanes, which DAG-based sequence methods cannot express, and the paper reports higher Landmark and Reachability F1 than prior models on nuScenes and Argoverse 2.

What carries the argument

The carrying mechanism is the chain of graph expansions: a serialization that makes the lane graph a fixed-order autoregressive target. At step $n$, the model outputs the new node's discretized position, the set of existing nodes it can come from ("from"), the set it can go to ("to"), and the Bézier control points for those directed centerlines; this corresponds to filling in the $n$-th row and column of the adjacency matrix and the matching entries of the geometric matrix. The nodes are ordered by depth-first search, and coordinates are quantized into discrete bins so the whole graph is one vocabulary over tokens. A transformer decoder with a bird's-eye-view encoder based on the LSS lifting scheme predicts that token chain under a standard cross-entropy objective, with node-position tokens weighted double.

What would settle it

Train SeqGrowGraph and the best prior baseline RNTR on the original, un-re-segmented nuScenes lane graphs and evaluate both on the original ground truth with the same Landmark and Reachability metrics; if SeqGrowGraph no longer beats the baseline, the reported state-of-the-art comes from the re-segmented task definition rather than from the chain-of-expansions formulation.

Watch

Extended reading notes

Core claim

The paper's central claim is that a lane graph can be incrementally grown as a sequence of node introductions, and that an autoregressive transformer trained on this sequence produces more complete and more topologically accurate lane graphs than detection-based or DAG-based alternatives. Formally, the lane graph is a directed graph $G=(V,E)$ whose vertices are intersections or key topological nodes and whose edges are centerlines. Each expansion step adds vertex $v_n=(x_n,y_n)$; the adjacency matrix $A$ grows from $n\times n$ to $(n+1)\times(n+1)$, with the new row encoding outgoing edges and the new column encoding incoming edges; the geometric matrix $M$ stores for each directed edge the middle control point of a quadratic Bézier curve. A depth-first ordering turns the graph into a token sequence, and the model predicts the sequence step by step. On the default and the overlap-free PON split of nuScenes, and on Argoverse 2, the paper reports top Landmark and Reachability F1 scores compared with prior methods, and it attributes the gain to keeping node geometry and topology in one growth process rather than generating points and edges separately.

Load-bearing premise

The reported advantage depends on training on a re-segmented centerline ground truth (continuous nodes removed, centerlines split at fixed lengths), while the baseline numbers appear to be produced on the original ground truth, and the paper never states which protocol generated Tables 1 and 2.

Editorial extensions

If this is right

  • Because the representation is not restricted to acyclic graphs, loops and bidirectional lanes can be generated directly, without the pre-processing and post-processing required by DAG-based methods.
  • Point geometry and connectivity are produced in the same token stream, so the model cannot emit a centerline without also saying which intersections it connects.
  • The depth-first ordering ablation indicates that node ordering is a real modeling choice; DFS gives the best Landmark and Reachability F1 in the paper's experiments.
  • The paper's centerline re-segmentation results suggest that lane-graph ground truth with continuous nodes removed and fixed-length segments is easier for a sequence model to learn, which the paper offers as a guideline for building future lane-graph datasets.

Reading between the lines

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

  • The paper leaves implicit that the same expansion serialization could be applied to other graph-structured map elements such as crosswalks, lane boundaries, or traffic-sign topology, since the only requirements are a node ordering and Bézier-geometry edges.
  • The re-segmentation protocol makes the headline comparison ambiguous: a fair state-of-the-art claim would require baselines retrained and evaluated on the same re-segmented ground truth, or SeqGrowGraph evaluated on the original ground truth; this is my inference, not something the paper states.
  • The DFS-versus-BFS ablation suggests that a learned or adaptive node-ordering policy might push the numbers further, but the paper only tests fixed orderings.
  • If the representation were paired with a large corpus of HD maps, the same chain-of-expansions target could serve as a pretraining objective for map-structure understanding; this is a natural extension the paper does not pursue.
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

3 major / 7 minor

Summary. SeqGrowGraph proposes to serialize a lane graph as a chain of node-insertion steps: each step appends a node's position, its incoming/outgoing Bézier control points, and the corresponding adjacency updates, with a depth-first ordering; a transformer decoder autoregressively predicts this sequence from BEV features. The paper reports Landmark and Reachability F1 on nuScenes (default and PON splits) and Argoverse 2, claiming state-of-the-art results, plus ablations on loss weighting and node ordering and a centerline re-segmentation study.

Significance. If the reported gains hold under the standard benchmark definition, the work is a useful contribution: it offers a representation that handles cycles and bidirectional lanes without DAG preprocessing, and the improvements over RNTR and LaneGraph2Seq are consistent across datasets. However, the significance currently hinges on an unspecified relation between the main tables and the re-segmentation protocol, and on single-run comparisons without code release.

major comments (3)
  1. [Section 4.5 / Tables 1 and 2] The protocol behind the headline numbers is not stated. Section 4.5 defines a new ground-truth construction for nuScenes: continuous nodes are removed, centerlines are re-split at 20/30/40 m, and the Landmark/Reachability metrics are recomputed in a 'Junction' variant that drops continuous nodes from both ground truth and predictions. The captions of Tables 1 and 2 and the comparison text in Section 4.4 do not say whether the SeqGrowGraph rows use this re-segmented graph and the Junction metric or the original RNTR graph with the original metric. This is load-bearing because Landmark F1 operates on the node set: changing the segmentation changes the number of landmarks, their positions, and the matching difficulty, while all baselines were trained and evaluated on the original RNTR annotations. The paper must state the exact protocol for every table and, ideally, report both variants. Table 4 does not resolve the ambiguity: its 'Original' row (Junction L-F 59.5, R-F 62.2) is not comparable to Table 1's default row (L-F 56.4, R-F 67.8), so the reader cannot infer which variant produced the SOTA claim.
  2. [Section 4.5 / Table 2] The re-segmentation discussion is confined to nuScenes (Figure 6); there is no statement about whether Argoverse 2 was re-segmented in the same way. If Table 2 uses a different graph definition from the nuScenes experiments, the method's consistency across datasets is not established, and the comparison with RNTR* and LaneGraph2Seq* on AV2 is unverifiable. The paper should specify the AV2 preprocessing and metric, or state explicitly that Table 2 uses the original annotation.
  3. [Section 4.4 / Tables 1 and 2] The reported results are single runs without error bars or significance tests, and no code is released. Given that the Landmark F1 advantage over the best baseline is 2.5–3.6 points and the paper tunes the loss weight and ordering on the same validation sets, the stability of the claimed improvement is not demonstrated. At minimum, report multiple seeds and standard deviations for the main tables.
minor comments (7)
  1. [Section 3.2] The definitions of F_n and T_n are written as sums over M(k,n) and M(n,k) for all k, but should be restricted to k with A(k,n)=1 (or A(n,k)=1); otherwise the notation suggests summing over empty entries.
  2. [Section 3.3] The heading 'Obejctive' should be 'Objective'.
  3. [Section 4.6] 'LaneGraph2Path' should be 'LaneGraph2Seq'.
  4. [Table 6] The headers 'BF S' and 'DF S' should be 'BFS' and 'DFS'.
  5. [Section 4.5] The phrase 'the metrics retrained after re-segmentation' is ungrammatical; it should read 'the metrics of models retrained after re-segmentation'.
  6. [Section 2.2] The statement about language models for autonomous driving is supported only by references [5] and [31], both from the same group; adding independent references would strengthen the contextual claim.
  7. [General] No code release or trained models are provided, which limits reproducibility; this should be acknowledged in the paper.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the graph-serialization derivation is self-contained; the §4.5 re-segmentation ambiguity is a benchmark-comparison concern, not a circular reduction.

full rationale

SeqGrowGraph's core construction is not defined from its target numbers. The paper serializes a lane graph as Sn = Sn-1 + (vn + Fn + Tn), with Fn/Tn collecting Bézier control points of edges to prior nodes, then trains a transformer with a standard cross-entropy MLE loss; no fitted scalar is renamed as a prediction and no result is derived from its own output. The DFS ordering and positional loss weight are selected by ablations (Tables 5-6), which is ordinary hyperparameter choice rather than circularity. The self-citations ([5], [30], [31]) appear in related-work context and are not load-bearing. The one substantive concern is §4.5: the paper removes 'continuous nodes' and re-segments centerlines at 20/30/40 m intervals, and its 'Junction' variant drops continuous nodes from ground truth and predictions, while Tables 1-2 do not state which ground-truth protocol produced the reported Landmark/Reachability numbers. If the headline gains were computed under the re-segmented protocol while baselines were evaluated on the original RNTR graph, the SOTA claim would be an apples-to-oranges comparison. That is a correctness/reproducibility risk, however, not a circular derivation: the re-segmentation is not an input whose value is recovered as the output, and no equation in the paper forces the reported F1 scores by construction. Accordingly the circularity score is low.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on the lane graph being a directed graph of junction nodes with Bézier centerlines, a lossless DFS-based serialization, and standard autoregressive training. The main added design choices are the point-loss weight (2), the 0.5 m token grid, and the re-segmentation length, none of which is justified by external data.

free parameters (3)
  • point position loss weight = 2
    Chosen from ablation in Table 5 over weights 1-4; the rest of the tokens use weight 1.
  • discretization resolution = 0.5 m
    Coordinates binned into 200 bins over a 100 m range (Section 4.2); no sensitivity analysis.
  • re-segmentation segment length = 20 m, 30 m, or 40 m (final choice not stated)
    Section 4.5 and Table 4 compare fixed-length splits; the paper does not say which split generates the training set for the main model.
assumptions (4)
  • domain assumption Lane graphs are directed graphs G=(V,E) where V are intersections/topological nodes and E are centerlines, with shape captured by quadratic Bézier control points.
    Section 3.1 defines the representation; this is a modeling choice, not a derived fact.
  • standard math The serialization Sn = S_{n-1} + (v_n + F_n + T_n) with a fixed DFS ordering is a lossless encoding of G.
    Section 3.2 defines the recursion; losslessness holds by construction if the ordering and the order of control points in F_n and T_n are fixed.
  • domain assumption Quadratic Bézier curves with one control point suffice to approximate centerline shapes after re-segmentation.
    Section 4.5 adds extra topological nodes to make the quadratic fit adequate, implying the original centerlines are not all quadratic.
  • domain assumption The 0.5 m discretization does not destroy topology.
    Section 4.2 maps coordinates to integer bins; no analysis of discretization error is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions." pith.science (2026). https://pith.science/paper/NXAWF3FF

@misc{pith2026250704822,
  author       = {Pith},
  title        = {Pith review of: SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NXAWF3FF}},
  note         = {Machine review of arXiv:2507.04822}
}
abstract

Accurate lane topology is essential for autonomous driving, yet traditional methods struggle to model the complex, non-linear structures-such as loops and bidirectional lanes-prevalent in real-world road structure. We present SeqGrowGraph, a novel framework that learns lane topology as a chain of graph expansions, inspired by human map-drawing processes. Representing the lane graph as a directed graph $G=(V,E)$, with intersections ($V$) and centerlines ($E$), SeqGrowGraph incrementally constructs this graph by introducing one vertex at a time. At each step, an adjacency matrix ($A$) expands from $n \times n$ to $(n+1) \times (n+1)$ to encode connectivity, while a geometric matrix ($M$) captures centerline shapes as quadratic B\'ezier curves. The graph is serialized into sequences, enabling a transformer model to autoregressively predict the chain of expansions, guided by a depth-first search ordering. Evaluated on nuScenes and Argoverse 2 datasets, SeqGrowGraph achieves state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2507.04822 by the authors.

Figure 1
Figure 1. Describing the lane graph step by step. Starting from an initial point, it is necessary to determine the location of the next point and establish whether there is a “from” or “to” relationship with the existing points. By repetitively carrying out this process, a complete description of the lane graph can be obtained. The abbreviation “pos” in the diagram stands for “position”. (BEV) space on a per-pixel basis but s… view at source ↗
Figure 2
Figure 2. Visualization comparison of lane graph results from different methods. This figure illustrates the inference results using the official models of TopoNet [10] and LaneGAP [15]. It can be observed that TopoNet, by detecting centerlines separately, results in discontinuities between the centerlines. In contrast, LaneGAP, which detects at the path level, tends to confuse multiple lanes and duplicate paths. Both methods… view at source ↗
Figure 3
Figure 3. Expansion process of the adjacency matrix. Each node corresponding to an intersections or key topological node and edges representing centerlines. At each step, we describe a new node along with the its connected centerlines. We focus only on the topological relationships between the newly introduced node and the existing nodes. To facilitate better formatting into a sequence, we use the new node as the subject to d… view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Model architecture and inference process. We employ a BEV-encoder to transform surrounding camera perspective view (PV) images into BEV features. Subsequently, a Transformer de￾coder generates tokens of the target sequence based on the BEV features and the beginning of…
Figure 6
Figure 6. Figure 6: Various definitions of centerline split have been consid [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: We visualize and compare the final results of various [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 31 canonical work pages

  1. [1]

    Curveformer: 3d lane detection by curve propagation with curve queries and attention

    Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In ICRA,

  2. [2]

    Ar- trackv2: Prompting autoregressive tracker where to look and how to describe

    Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Ar- trackv2: Prompting autoregressive tracker where to look and how to describe. In CVPR, pages 19048–19057, 2024. 1

  3. [3]

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

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 5

  4. [4]

    Structured bird’s-eye-view traffic scene un- derstanding from onboard images

    Yigit Baran Can, Alexander Liniger, Danda Pani Paudel, and Luc Van Gool. Structured bird’s-eye-view traffic scene un- derstanding from onboard images. In ICCV, 2021. 1, 3, 6

  5. [5]

    Driving by the rules: A benchmark for integrating traffic sign regulations into vectorized hd map, 2025

    Xinyuan Chang, Maixuan Xue, Xinran Liu, Zheng Pan, and Xing Wei. Driving by the rules: A benchmark for integrating traffic sign regulations into vectorized hd map, 2025. 3

  6. [6]

    Pix2seq: A language modeling framework for object detection

    Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Ge- offrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. 1

  7. [7]

    Can llms effectively leverage graph structural information through prompts, and why? Trans

    Jin Huang, Xingjian Zhang, Qiaozhu Mei, and Jiaqi Ma. Can llms effectively leverage graph structural information through prompts, and why? Trans. Mach. Learn. Res., 2024. 3

  8. [8]

    Repvf: A unified vector fields representation for multi-task 3d perception

    Chunliang Li, Wencheng Han, Jun Yin, Sanyuan Zhao, and Jianbing Shen. Repvf: A unified vector fields representation for multi-task 3d perception. In ECCV, 2024. 1

Show all 33 references
  1. [9]

    Hdmapnet: An online HD map construction and evaluation framework

    Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In ICRA, 2022. 1, 2

  2. [10]

    Graph- based topology reasoning for driving scenes, 2023

    Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xi- angwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chun- jing Xu, Junchi Yan, Ping Luo, and Hongyang Li. Graph- based topology reasoning for driving scenes, 2023. 1, 2, 3, 6, 7

  3. [11]

    Grou- plane: End-to-end 3d lane detection with channel-wise group- ing

    Zhuoling Li, Chunrui Han, Zheng Ge, Jinrong Yang, En Yu, Haoqian Wang, Xiangyu Zhang, and Hengshuang Zhao. Grou- plane: End-to-end 3d lane detection with channel-wise group- ing. Robotics Autom. Lett., 2024. 2

  4. [12]

    Autore- gressive sequential pretraining for visual tracking

    Shiyi Liang, Yifan Bai, Yihong Gong, and Xing Wei. Autore- gressive sequential pretraining for visual tracking. In CVPR, pages 7254–7264, 2025. 1

  5. [13]

    Maptr: Structured modeling and learning for online vectorized HD map construction

    Bencheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 2

  6. [14]

    Map- trv2: An end-to-end framework for online vectorized HD map construction

    Bencheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Map- trv2: An end-to-end framework for online vectorized HD map construction. CoRR, 2023. 2

  7. [15]

    Lane graph as path: Continuity-preserving path-wise model- ing for online lane graph construction

    Bencheng Liao, Shaoyu Chen, Bo Jiang, Tianheng Cheng, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Lane graph as path: Continuity-preserving path-wise model- ing for online lane graph construction. In ECCV, 2024. 1, 2, 3, 6, 7

  8. [16]

    Vectormapnet: End-to-end vectorized HD map learning

    Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized HD map learning. In ICML, 2023. 2

  9. [17]

    Translating images to road network: A non-autoregressive sequence to-sequence approach

    Jiachen Lu, Hongyang Li, Renyuan Peng, Feng Wen, Xinyue Cai, Wei Zhang, Hang Xu, and Li Zhang. Translating images to road network: A non-autoregressive sequence to-sequence approach. In ICCV, 2023. 2, 3, 5, 6, 8

  10. [18]

    Luo, Xinshuo Weng, Yan Wang, Shuang Wu, Jie Li, Kilian Q

    Katie Z. Luo, Xinshuo Weng, Yan Wang, Shuang Wu, Jie Li, Kilian Q. Weinberger, Yue Wang, and Marco Pavone. Augmenting lane perception and topology understanding with standard definition navigation maps. In ICRA, 2024. 1

  11. [19]

    LATR: 3d lane detection from monocular images with transformer

    Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. LATR: 3d lane detection from monocular images with transformer. In ICCV, 2023. 3

  12. [20]

    GNN-LM: language modeling based on global contexts via GNN

    Yuxian Meng, Shi Zong, Xiaoya Li, Xiaofei Sun, Tianwei Zhang, Fei Wu, and Jiwei Li. GNN-LM: language modeling based on global contexts via GNN. In ICLR, 2022. 3

  13. [21]

    Lanegraph2seq: Lane topology extraction with language model via vertex-edge encoding and connectivity enhancement

    Renyuan Peng, Xinyue Cai, Hang Xu, Jiachen Lu, Feng Wen, Wei Zhang, and Li Zhang. Lanegraph2seq: Lane topology extraction with language model via vertex-edge encoding and connectivity enhancement. In AAAI, 2024. 2, 3, 5, 6, 8

  14. [22]

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

    Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 5, 6

  15. [23]

    Lanecpp: Continuous 3d lane detection using physi- cal priors

    Maximilian Pittner, Joel Janai, and Alexandru Paul Con- durache. Lanecpp: Continuous 3d lane detection using physi- cal priors. In CVPR, 2024. 1, 2

  16. [24]

    Predicting semantic map representations from images using pyramid occupancy networks

    Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In CVPR, 2020. 5, 6

  17. [25]

    Openlane-v2: A topology rea- soning benchmark for unified 3d HD mapping

    Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. Openlane-v2: A topology rea- soning benchmark for unified 3d HD mapping. In NeurIPS,

  18. [26]

    Autoregressive visual tracking

    Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, pages 9697– 9706, 2023. 1

  19. [27]

    Argoverse 2: Next generation datasets for self-driving perception and fore- casting

    Benjamin Wilson, William Qi, Tanmay Agarwal, John Lam- bert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Rat- nesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception an...

  20. [28]

    Sparse point guided 3d lane detection

    Chengtang Yao, Lidong Yu, Yuwei Wu, and Yunde Jia. Sparse point guided 3d lane detection. In ICCV, 2023. 2

  21. [29]

    Language is all a graph needs

    Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang. Language is all a graph needs. In Findings of the Association for Computational Linguistics: EACL 2024, St. Julian’s, Malta, March 17-22, 2024, 2024. 3

  22. [30]

    Driving with prior maps: Unified vector prior encoding for autonomous vehicle mapping

    Shuang Zeng, Xinyuan Chang, Xinran Liu, Zheng Pan, and Xing Wei. Driving with prior maps: Unified vector prior encoding for autonomous vehicle mapping. arXiv preprint arXiv:2409.05352, 2024. 2

  23. [31]

    Futuresightdrive: 9 Thinking visually with spatio-temporal cot for autonomous driving, 2025

    Shuang Zeng, Xinyuan Chang, Mengwei Xie, Xinran Liu, Yi- fan Bai, Zheng Pan, Mu Xu, and Xing Wei. Futuresightdrive: 9 Thinking visually with spatio-temporal cot for autonomous driving, 2025. 3

  24. [32]

    Pvalane: Prior-guided 3d lane detection with view- agnostic feature alignment

    Zewen Zheng, Xuemin Zhang, Yongqiang Mou, Xiang Gao, Chengxin Li, Guoheng Huang, Chi-Man Pun, and Xiaochen Yuan. Pvalane: Prior-guided 3d lane detection with view- agnostic feature alignment. In AAAI, 2024. 1, 3

  25. [33]

    Himap: Hybrid representa- tion learning for end-to-end vectorized HD map construction

    Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representa- tion learning for end-to-end vectorized HD map construction. In CVPR, 2024. 2 10

Pith tools

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