Pith. sign in

REVIEW 3 major objections 6 minor 55 references

Reusing Attention for One-stage Lane Topology Understanding

T0 review · 3 major / 6 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 Lane and traffic-element topology can be predicted in the same pass as detection by reusing intermediate transformer attention, improving OpenLane-V2 scores while cutting inference time by 17%.

desk verdict A competent, incremental one-stage adaptation of EGTR's attention reuse for lane topology, where the real accuracy gain comes from map-to-mapless distillation rather than the architecture itself. read the letter →

arxiv 2507.17617 v1 pith:IOQGF23H submitted 2025-07-23 cs.CV

Yang Li , Zongzheng Zhang , Xuchong Qiu , Xinrun Li , Ziming Liu , Leichen Wang , Ruikai Li , Zhenxin Zhu
show 5 more authors
Huan-ang Gao Xiaojian Lin Zhiyong Cui Hang Zhao Hao Zhao
This is my paper
classification cs.CV
keywords lanetopologyunderstandingone-stagedetectionattentionreusetransformerdecodertrafficelementscenterlinesknowledgedistillationOpenLane-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 tries to establish that lane and traffic-element topology can be predicted in the same forward pass as detection, without a separate graph construction stage. It proposes reusing the intermediate self-attention queries and keys of two transformer decoders—one for traffic elements in perspective view and one for lane centerlines in bird's-eye view—and feeding their pairwise concatenations through gated sums and MLPs to output lane-lane and lane-traffic-element relation matrices. On OpenLane-V2 subset-A the resulting one-stage student reaches 40.6 OLS without SD maps and the SD-map-based teacher reaches 44.3 OLS, both above their two-stage baselines while running 17% faster with fewer parameters. The paper further claims that distilling BEV features from an SD-map-based teacher into a mapless student transfers part of the map's benefit, improving the student's OLS from 39.9 to 40.6.

What carries the argument

The load-bearing mechanism is cross-decoder attention reuse: intermediate queries $Q^l_{TE}, Q^l_{CL}$ and keys $K^l_{TE}, K^l_{CL}$ are drawn from each self-attention layer of the two decoders, linearly projected, concatenated pairwise into relation resources $R^{1:L}$, and combined with a last-layer relation resource $R^z$ (Eqs. (5)-(6)); the task-relevant sub-blocks are processed by gated sums and MLPs. The companion mechanism is map-to-mapless knowledge distillation, which aligns student and teacher BEV features with an MSE loss so that a camera-only student inherits the relational structure learned with SD-map input.

What would settle it

Train the same one-stage model but replace the intermediate queries and keys with random noise before the pairwise-concatenation step, keeping the detection heads unchanged; if topology scores on OpenLane-V2 subset-A stay at the same level, the reported gains do not come from attention reuse. A cheaper test is to ablate the intermediate-layer terms and keep only the last-layer decoder outputs in $R^z$: if OLS does not drop, intermediate self-attention resources are unnecessary.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the self-attention resources of object-detection transformers already contain the relational information needed for lane topology, so topology prediction can be folded into the detection decoders. The authors extract queries and keys from every self-attention layer of the traffic-element and lane-centerline decoders, linearly project them, concatenate them pairwise into a relation tensor $R$ (Eqs. (3)-(6)), and select the blocks $R_{TECL}$ and $R_{CLCL}$ that correspond to lane-traffic-element and lane-lane pairs; gated sums and MLPs then produce the topology matrices. Because detection and topology share the same decoder features and are optimized jointly, the two-stage error-propagation path through a separately built graph is removed. As a corollary, the paper shows that a teacher model using SD maps can be converted to this one-stage design and then used to distill mapless students, making the SD-map's accuracy benefit partly available without requiring the map at inference.

Load-bearing premise

The method assumes the self-attention queries and keys inside the detection decoders already encode enough relational structure between traffic elements and lanes that a lightweight head can read topology off them; if the decoders' attention is only about local appearance, the one-stage topology prediction has no signal to exploit.

Editorial extensions

If this is right

  • Topology reasoning no longer needs a separately constructed graph or a GNN, removing the extra computation and parameters those modules add; the paper reports a 17% inference-time reduction and about 13 million fewer parameters versus its baselines.
  • Detection and topology are optimized jointly, so errors from a frozen detection stage cannot propagate into relation prediction the way they do in two-stage pipelines.
  • The one-stage design transfers to both settings with and without SD maps: the teacher improves over SMERF on the OpenLane-V2 subset-A OLS metric, and the distilled student improves over TopoNet.
  • Distillation from map-based teachers to mapless students works through BEV-feature alignment, narrowing the accuracy gap caused by missing map input at no additional inference cost.
  • Using extra feature interactions beyond the task-relevant topology blocks hurts accuracy, indicating that selective relation features matter more than broader mixing.

Reading between the lines

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

  • The same attention-reuse construction could apply to other paired detection-and-relation problems, such as vehicle-to-lane or vehicle-to-vehicle interaction graphs, since the mechanism is not specific to traffic elements or centerlines.
  • The degradation from extra feature interactions suggests the relational signal is concentrated in the selected blocks; a follow-up could test whether replacing the gated-sum combiner with learned attention pooling preserves that concentration.
  • The distillation result suggests SD maps function mainly as a feature regularizer rather than an indispensable input; if so, ensembling map-based teachers or adding temporal context could push mapless students closer to the teacher's 44.3 OLS.
  • The reported 17% speed gain excludes feature extraction time, so the end-to-end latency benefit for a full autonomous-driving perception stack may differ from the decoder-level comparison.
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 / 6 minor

Summary. The paper proposes a one-stage lane topology understanding architecture in which intermediate self-attention queries and keys from separate traffic-element (TE) and lane-centerline (CL) transformer decoders are reused, via linear projection, pairwise concatenation, and gated sums, to predict CL-CL and TE-CL topology simultaneously with detection. A second contribution is a knowledge-distillation scheme that transfers BEV features from an SD-map-based teacher to an SD-map-free student. Experiments on OpenLane-V2 subset-A/B report OLS 44.3 (teacher) vs 43.0 for SMERF and OLS 40.6 (student) vs 39.8 for TopoNet, together with a claimed 17% speed advantage and lower parameter counts. Code is released.

Significance. The attention-reuse topology head is a plausible lightweight alternative to explicit graph networks, and the map-to-mapless distillation is a practically motivated transfer idea. The evaluation uses the standard OpenLane-V2 benchmark with backbones matched to the baselines, includes an ablation of the distillation component, and reports parameter counts; these are clear strengths. However, two issues affect the central claims. First, the speed advantage is measured after excluding feature extraction, which dominates runtime in this pipeline, so the reported 17% is not established as an end-to-end speed gain. Second, on subset-A the SD-map-free student without distillation does not beat TopoNet in OLS (39.9 vs 39.8, Table IV vs Table I), so the headline accuracy gain in the map-free setting is driven by distillation rather than by the one-stage design per se. These points need to be addressed before the contribution can be assessed as stated.

major comments (3)
  1. [IV-B and Table II] The speed comparison excludes feature extraction time, as stated in Sec. IV-B. Since the image backbone, FPN, and view transformer are shared with TopoNet and SMERF, the +17% in Table II reflects only decoder plus topology-head cost. For a multi-view driving pipeline, the excluded components typically dominate latency, so the end-to-end speed advantage is likely much smaller and could be within run-to-run noise. Please report full-system inference time on the same GPU and batch for all methods, and state unambiguously whether the table reports throughput or latency. This is load-bearing because improving inference speed is a central claim in the abstract and introduction.
  2. [Tables I and IV] On OpenLane-V2 subset-A, the SD-map-free student without distillation scores OLS 39.9 (Table IV), essentially equal to TopoNet's 39.8 (Table I). The reported 40.6 OLS for the student is obtained only after BEV-feature distillation from the SD-map-based teacher. Thus the abstract's claim that the one-stage architecture improves accuracy is not supported on the primary benchmark in the SD-map-free setting; that improvement is attributable to distillation. Please qualify the accuracy claim, for example by giving the subset-B result (Table III) a more prominent role or by adding a benchmark where the one-stage design alone shows a clear gain.
  3. [III-B, Eq. (6)] The definition of Rz is dimensionally unclear: it is stated to be in R^{(N_TE+N_CL) x (N_TE+N_CL) x L x d} while being formed from 'the last-layer decoder outputs,' which are single tensors of shape N x d. The stacking operation in Eq. (6) also does not specify the concatenation axis. Please clarify whether Rz has an L dimension (for instance, by collecting the output after each decoder layer) and give the final shape of R. Without this clarification, the topology head cannot be reproduced from the text alone.
minor comments (6)
  1. [Abstract] The abstract contains typos: 'toplology' should be 'topology' and 'models that utilize standard definition (SD) maps to those operates without using SD maps' should be 'models that operate without using SD maps.'
  2. [IV-C] The text refers to 'STST [4]' but Table I and the reference list use 'STSU [4]'; please correct this inconsistency.
  3. [III-B] The phrase 'passed throgh the centerline transformer' should read 'passed through the centerline transformer.'
  4. [Table II] The column header 'Inference Speed ↑' is ambiguous; if it denotes relative throughput, state that explicitly in the caption to avoid confusion with the '17% reduction in inference time' wording in the introduction.
  5. [IV-E.2] In the feature-interaction ablation, 'the complete matrix R' is not precisely defined in relation to the task-relevant blocks R_TECL and R_CLCL; please specify which blocks of R are added in that experiment and how the conv+pool then selects the CL-CL output.
  6. [Tables I, III, IV] The reported OLS differences are small (0.8 and 1.3 points), and no variance or multiple-seed results are given; please report at least one additional seed or a significance test for the main comparisons.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the attention-reuse topology head is an architectural design trained and evaluated on held-out OpenLane-V2 labels, with no prediction reducing to a fitted input or self-citation chain.

full rationale

The paper's central derivation is self-contained in the sense required here: the topology features are formed from intermediate queries and keys (Eqs. 1-8), and the resulting relation matrices are supervised by the same detection and topology losses used in TopoNet, evaluated on the public OpenLane-V2 benchmark. The evaluation targets are ground-truth annotations, not quantities constructed from the model's own outputs. The distillation objective (Eq. 9) compares student and teacher BEV features via MSE and uses soft labels from a frozen teacher; this is a standard training procedure, and the teacher is itself trained on the same dataset with additional SD-map inputs, not on the test set. No uniqueness theorem or prior self-citation is invoked to make the architectural choice forced, and the attention-reuse mechanism is explicitly contrasted with EGTR rather than presented as an independent external result. The only notable concern is that the reported 17% speed advantage excludes feature extraction time, which is a measurement-scope issue for the speed claim, not a circularity issue. Accordingly, no circular step meeting the required evidentiary standard was found.

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

The architecture introduces no new physical entities. Its load-bearing assumptions are that detection-decoder attention features encode topology-relevant relations, that OpenLane-V2 annotations are reliable, and that the hand-crafted fusion and distillation designs are suitable.

free parameters (3)
  • Distillation loss weight = Not reported (presumed 1.0)
    The student objective in Sec. III-C adds L_BEV (Eq. 9) to the task losses without stating a weighting coefficient; this balance affects the student's accuracy gain (Table IV).
  • Decoder layer count and query numbers = Inherited from TopoNet/SMERF
    Section IV-B states the counts are the same as baselines; specific integer values are not given, so the relation matrix size R(NTE+NCL)x... cannot be reproduced from the text alone.
  • Projection dimension d/2 = Half of the transformer dimension
    Eq. (1)-(2) project queries and keys to d/2 before concatenation; this choice is inherited from EGTR and not ablated.
assumptions (4)
  • domain assumption DETR-style self-attention in detection decoders encodes object relations useful for topology prediction
    The whole method in Sec. III-B assumes that intermediate queries and keys from TE and CL decoders carry relational signal; this is introduced without theoretical justification.
  • domain assumption OpenLane-V2 annotations provide reliable ground truth for lane, traffic element, and topology
    The training and evaluation in Sec. IV-A rely on OpenLane-V2 labels to define the target and measure performance.
  • domain assumption BEV features from the view transform in [7] are sufficient for lane detection and topology
    The CL decoder operates only on BEV features built from the TopoNet view transform; no ablations on feature quality are provided.
  • ad hoc to paper Linear projection and pairwise concatenation preserve the relational structure in the attention features
    The fusion design in Eq. (1)-(8) is a hand-crafted choice imported from EGTR, not derived from first principles or proven optimal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reusing Attention for One-stage Lane Topology Understanding." pith.science (2026). https://pith.science/paper/IOQGF23H

@misc{pith2026250717617,
  author       = {Pith},
  title        = {Pith review of: Reusing Attention for One-stage Lane Topology Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOQGF23H}},
  note         = {Machine review of arXiv:2507.17617}
}
read the original abstract

Understanding lane toplogy relationships accurately is critical for safe autonomous driving. However, existing two-stage methods suffer from inefficiencies due to error propagations and increased computational overheads. To address these challenges, we propose a one-stage architecture that simultaneously predicts traffic elements, lane centerlines and topology relationship, improving both the accuracy and inference speed of lane topology understanding for autonomous driving. Our key innovation lies in reusing intermediate attention resources within distinct transformer decoders. This approach effectively leverages the inherent relational knowledge within the element detection module to enable the modeling of topology relationships among traffic elements and lanes without requiring additional computationally expensive graph networks. Furthermore, we are the first to demonstrate that knowledge can be distilled from models that utilize standard definition (SD) maps to those operates without using SD maps, enabling superior performance even in the absence of SD maps. Extensive experiments on the OpenLane-V2 dataset show that our approach outperforms baseline methods in both accuracy and efficiency, achieving superior results in lane detection, traffic element identification, and topology reasoning. Our code is available at https://github.com/Yang-Li-2000/one-stage.git.

Figures

Figures reproduced from arXiv: 2507.17617 by the authors.

Figure 1
Figure 1. (a) Two-stage methods typically use multi-view images and auxiliary SD maps as input for the detection-prediction [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall Architecture. We propose a novel one-stage method where intermediate queries (QT E and QCL) and keys (KT E and KCL) are extracted from each self-attention layer within the traffic elements (TE) and lane centerlines (CL) decoders, and are subsequently used for topology reasoning. Concurrently, traffic elements and centerlines are predicted using the final-layer decoder outputs. We further introduce a teacher-… view at source ↗
Figure 3
Figure 3. Topology Feature Processing. Intermediate queries and keys extracted from each self-attention layer within the traffic elements (TE) and centerlines (CL) decoders are projected using linear layers and concatenated. Outputs from the last layers of the two decoders are also projected using linear layers. After that, projected features are pairwise concatenated. Finally, task-relevant projected features are selected an… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative Comparisons between TopoNet [7] and our student network. Left (Multi-View Inputs): Visualization of corresponding multi-view inputs. Middle (CL and CLCL Predictions): Purple indicates false positives, while blue denotes true positives. Right (TE and TECL Pr…
Figure 5
Figure 5. Figure 5: Qualitative Comparisons between TopoNet [7] and our student network. Left (Multi-View Inputs): Visualization of corresponding multi-view inputs. Middle (CL and CLCL Predictions): Purple indicates false positives, while blue denotes true positives. Right (TE and TECL Pr…
Figure 6
Figure 6. Figure 6: Qualitative Comparisons between our network with interactions and our teacher network. Left (CL and CLCL Predictions): Purple indicates false positives, while blue denotes true positives. Right (TE and TECL Predictions): Green represents true positives, whereas red sig…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 46 canonical work pages

  1. [1]

    Openlane-v2: A topology reasoning bench- mark for unified 3d hd mapping,

    H. Wang, T. Li, Y . Li, L. Chen, C. Sima, Z. Liu, B. Wang, P. Jia, Y . Wang, S. Jiang et al. , “Openlane-v2: A topology reasoning bench- mark for unified 3d hd mapping,” Advances in Neural Information Processing Systems, vol. 36, pp. 18 873–18 884, 2023

    work page 2023

  2. [2]

    Unsuper- vised road anomaly detection with language anchors,

    B. Tian, M. Liu, H.-a. Gao, P. Li, H. Zhao, and G. Zhou, “Unsuper- vised road anomaly detection with language anchors,” in 2023 IEEE international conference on robotics and automation (ICRA) . IEEE, 2023, pp. 7778–7785

    work page 2023

  3. [3]

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

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

    work page 2024

  4. [4]

    Structured bird’s-eye-view traffic scene understanding from onboard images,

    Y . B. Can, A. Liniger, D. P. Paudel, and L. Van Gool, “Structured bird’s-eye-view traffic scene understanding from onboard images,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 15 661–15 670

    work page 2021

  5. [5]

    Vectormapnet: End-to-end vectorized hd map learning,

    Y . Liu, T. Yuan, Y . Wang, Y . Wang, and H. Zhao, “Vectormapnet: End-to-end vectorized hd map learning,” in International Conference on Machine Learning . PMLR, 2023, pp. 22 352–22 369

    work page 2023

  6. [6]

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

    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,” arXiv preprint arXiv:2208.14437 , 2022

    arXiv 2022

  7. [7]

    Graph-based topology reasoning for driving scenes,

    T. Li, L. Chen, H. Wang, Y . Li, J. Yang, X. Geng, S. Jiang, Y . Wang, H. Xu, C. Xu et al. , “Graph-based topology reasoning for driving scenes,” arXiv preprint arXiv:2304.05277 , 2023

    arXiv 2023

  8. [8]

    Egtr: Extracting graph from transformer for scene graph generation,

    J. Im, J. Nam, N. Park, H. Lee, and S. Park, “Egtr: Extracting graph from transformer for scene graph generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 24 229–24 238

    work page 2024

Show all 55 references
  1. [9]

    Building lane-level maps from aerial images,

    J. Yao, X. Pan, T. Wu, and X. Zhang, “Building lane-level maps from aerial images,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 3890–3894

    2024

  2. [10]

    Automatic con- struction of lane-level hd maps for urban scenes,

    Y . Zhou, Y . Takeda, M. Tomizuka, and W. Zhan, “Automatic con- struction of lane-level hd maps for urban scenes,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2021, pp. 6649–6656

    2021

  3. [11]

    Lane graph estimation for scene understanding in urban driving,

    J. Z ¨urn, J. Vertens, and W. Burgard, “Lane graph estimation for scene understanding in urban driving,” IEEE Robotics and Automation Letters, vol. 6, no. 4, pp. 8615–8622, 2021

    2021

  4. [12]

    Tod3cap: Towards 3d dense captioning in outdoor scenes,

    B. Jin, Y . Zheng, P. Li, W. Li, Y . Zheng, S. Hu, X. Liu, J. Zhu, Z. Yan, H. Sun et al. , “Tod3cap: Towards 3d dense captioning in outdoor scenes,” in European Conference on Computer Vision. Springer, 2024, pp. 367–384

    2024

  5. [13]

    Topology-guided road graph extraction from remote sensing images,

    Y . Zao, Z. Zou, and Z. Shi, “Topology-guided road graph extraction from remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 62, pp. 1–14, 2023

    2023

  6. [14]

    Improving online lane graph extraction by object-lane clustering,

    Y . B. Can, A. Liniger, D. P. Paudel, and L. Van Gool, “Improving online lane graph extraction by object-lane clustering,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 8591–8601

    2023

  7. [15]

    Lane graph as path: Continuity-preserving path-wise modeling for online lane graph construction,

    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 European Conference on Computer Vision . Springer, 2024, pp. 334–351

    2024

  8. [16]

    Centerlinedet: Cen- terline graph detection for road lanes with vehicle-mounted sensors by transformer for hd map generation,

    Z. Xu, Y . Liu, Y . Sun, M. Liu, and L. Wang, “Centerlinedet: Cen- terline graph detection for road lanes with vehicle-mounted sensors by transformer for hd map generation,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 3553–3559

    2023

  9. [17]

    Topomlp: An simple yet strong pipeline for driving topology reasoning,

    D. Wu, J. Chang, F. Jia, Y . Liu, T. Wang, and J. Shen, “Topomlp: An simple yet strong pipeline for driving topology reasoning,” arXiv preprint arXiv:2310.06753, 2023

    2023 arXiv

  10. [18]

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

    Z. Zhang, X. Li, S. Zou, G. Chi, S. Li, X. Qiu, G. Wang, G. Zheng, L. Wang, H. Zhao et al., “Chameleon: Fast-slow neuro-symbolic lane topology extraction,” arXiv preprint arXiv:2503.07485 , 2025

    2025 arXiv

  11. [19]

    Delving into shape-aware zero-shot semantic segmentation,

    X. Liu, B. Tian, Z. Wang, R. Wang, K. Sheng, B. Zhang, H. Zhao, and G. Zhou, “Delving into shape-aware zero-shot semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2999–3009

    2023

  12. [20]

    Scene graph generation by iterative message passing,

    D. Xu, Y . Zhu, C. B. Choy, and L. Fei-Fei, “Scene graph generation by iterative message passing,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 5410–5419

    2017

  13. [21]

    Graph r-cnn for scene graph generation,

    J. Yang, J. Lu, S. Lee, D. Batra, and D. Parikh, “Graph r-cnn for scene graph generation,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 670–685

    2018

  14. [22]

    Unbiased scene graph generation from biased training,

    K. Tang, Y . Niu, J. Huang, J. Shi, and H. Zhang, “Unbiased scene graph generation from biased training,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 3716–3725

    2020

  15. [23]

    Pq-transformer: Jointly parsing 3d objects and layouts from point clouds,

    X. Chen, H. Zhao, G. Zhou, and Y .-Q. Zhang, “Pq-transformer: Jointly parsing 3d objects and layouts from point clouds,” IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2519–2526, 2022

    2022

  16. [24]

    Understanding embodied reference with touch-line transformer

    Y . Li, X. Chen, H. Zhao, J. Gong, G. Zhou, F. Rossano, and Y . Zhu, “Understanding embodied reference with touch-line transformer.” in ICLR, 2023

    2023

  17. [25]

    Distance-aware occlusion detection with focused attention,

    Y . Li, Y . Tu, X. Chen, H. Zhao, and G. Zhou, “Distance-aware occlusion detection with focused attention,” IEEE Transactions on Image Processing, vol. 31, pp. 5661–5676, 2022

    2022

  18. [26]

    Cerberus transformer: Joint semantic, affordance and attribute parsing,

    X. Chen, T. Liu, H. Zhao, G. Zhou, and Y .-Q. Zhang, “Cerberus transformer: Joint semantic, affordance and attribute parsing,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 19 649–19 658

    2022

  19. [27]

    From semi-supervised to omni-supervised room layout estimation using point clouds,

    H.-a. Gao, B. Tian, P. Li, X. Chen, H. Zhao, G. Zhou, Y . Chen, and H. Zha, “From semi-supervised to omni-supervised room layout estimation using point clouds,” arXiv preprint arXiv:2301.13865 , 2023

    2023 arXiv

  20. [28]

    Toist: Task oriented instance segmentation transformer with noun- pronoun distillation,

    P. Li, B. Tian, Y . Shi, X. Chen, H. Zhao, G. Zhou, and Y .-Q. Zhang, “Toist: Task oriented instance segmentation transformer with noun- pronoun distillation,” Advances in Neural Information Processing Systems, vol. 35, pp. 17 597–17 611, 2022

    2022

  21. [29]

    Visual relationship detection with language priors,

    C. Lu, R. Krishna, M. Bernstein, and L. Fei-Fei, “Visual relationship detection with language priors,” in Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14 . Springer, 2016, pp. 852–869

    2016

  22. [30]

    Neural motifs: Scene graph parsing with global context,

    R. Zellers, M. Yatskar, S. Thomson, and Y . Choi, “Neural motifs: Scene graph parsing with global context,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 5831–5840

    2018

  23. [31]

    Linknet: Relational em- bedding for scene graph,

    S. Woo, D. Kim, D. Cho, and I. S. Kweon, “Linknet: Relational em- bedding for scene graph,” Advances in neural information processing systems, vol. 31, 2018

    2018

  24. [32]

    Learning to compose dynamic tree structures for visual contexts,

    K. Tang, H. Zhang, B. Wu, W. Luo, and W. Liu, “Learning to compose dynamic tree structures for visual contexts,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 6619–6628

    2019

  25. [33]

    Relation transformer network,

    R. Koner, S. Shit, and V . Tresp, “Relation transformer network,” ECCV, 2020

    2020

  26. [34]

    Gps-net: Graph property sensing network for scene graph generation,

    X. Lin, C. Ding, J. Zeng, and D. Tao, “Gps-net: Graph property sensing network for scene graph generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 3746–3753

    2020

  27. [35]

    Context-aware scene graph generation with seq2seq transformers,

    Y . Lu, H. Rai, J. Chang, B. Knyazev, G. Yu, S. Shekhar, G. W. Taylor, and M. V olkovs, “Context-aware scene graph generation with seq2seq transformers,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 15 931–15 941

    2021

  28. [36]

    Bgt-net: Bidirectional gru transformer network for scene graph generation,

    N. Dhingra, F. Ritter, and A. Kunz, “Bgt-net: Bidirectional gru transformer network for scene graph generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 2150–2159

    2021

  29. [37]

    Bipartite graph network with adaptive message passing for unbiased scene graph generation,

    R. Li, S. Zhang, B. Wan, and X. He, “Bipartite graph network with adaptive message passing for unbiased scene graph generation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 11 109–11 119

    2021

  30. [38]

    Environment-invariant curriculum rela- tion learning for fine-grained scene graph generation,

    Y . Min, A. Wu, and C. Deng, “Environment-invariant curriculum rela- tion learning for fine-grained scene graph generation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 13 296–13 307

    2023

  31. [39]

    Vision relation transformer for unbiased scene graph generation,

    G. Sudhakaran, D. S. Dhami, K. Kersting, and S. Roth, “Vision relation transformer for unbiased scene graph generation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 21 882–21 893

    2023

  32. [40]

    Composi- tional feature augmentation for unbiased scene graph generation,

    L. Li, G. Chen, J. Xiao, Y . Yang, C. Wang, and L. Chen, “Composi- tional feature augmentation for unbiased scene graph generation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 21 685–21 695

    2023

  33. [41]

    Faster r-cnn: Towards real- time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real- time object detection with region proposal networks,” Advances in neural information processing systems , vol. 28, 2015

    2015

  34. [42]

    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,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 779– 788

    2016

  35. [43]

    Pixels to graphs by associative embedding,

    A. Newell and J. Deng, “Pixels to graphs by associative embedding,” Advances in neural information processing systems , vol. 30, 2017

    2017

  36. [44]

    Fully convolutional scene graph generation,

    H. Liu, N. Yan, M. Mortazavi, and B. Bhanu, “Fully convolutional scene graph generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 11 546–11 556

    2021

  37. [45]

    Structured sparse r-cnn for direct scene graph generation,

    Y . Teng and L. Wang, “Structured sparse r-cnn for direct scene graph generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 19 437–19 446

    2022

  38. [46]

    Relationformer: A unified framework for image-to-graph generation,

    S. Shit, R. Koner, B. Wittmann, J. Paetzold, I. Ezhov, H. Li, J. Pan, S. Sharifzadeh, G. Kaissis, V . Tresp et al., “Relationformer: A unified framework for image-to-graph generation,” in European Conference on Computer Vision . Springer, 2022, pp. 422–439

    2022

  39. [47]

    Iterative scene graph generation,

    S. Khandelwal and L. Sigal, “Iterative scene graph generation,” Ad- vances in Neural Information Processing Systems , vol. 35, pp. 24 295– 24 308, 2022

    2022

  40. [48]

    Reltr: Relation transformer for scene graph generation,

    Y . Cong, M. Y . Yang, and B. Rosenhahn, “Reltr: Relation transformer for scene graph generation,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 9, pp. 11 169–11 183, 2023

    2023

  41. [49]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European conference on computer vision . Springer, 2020, pp. 213– 229

    2020

  42. [50]

    Sgtr: End-to-end scene graph generation with transformer,

    R. Li, S. Zhang, and X. He, “Sgtr: End-to-end scene graph generation with transformer,” in proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 19 486–19 496

    2022

  43. [51]

    Uniq: Unified decoder with task-specific queries for efficient scene graph generation,

    X. Liao, W. Wei, D. Chen, and Y . Fu, “Uniq: Unified decoder with task-specific queries for efficient scene graph generation,” in Proceed- ings of the 32nd ACM International Conference on Multimedia , 2024, pp. 8815–8824

    2024

  44. [52]

    Single-stage visual relationship learning using conditional queries,

    A. Desai, T.-Y . Wu, S. Tripathi, and N. Vasconcelos, “Single-stage visual relationship learning using conditional queries,” Advances in Neural Information Processing Systems , vol. 35, pp. 13 064–13 077, 2022

    2022

  45. [53]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 2117–2125

    2017

  46. [54]

    Monoocc: Digging into monocular semantic occupancy prediction,

    Y . Zheng, X. Li, P. Li, Y . Zheng, B. Jin, C. Zhong, X. Long, H. Zhao, and Q. Zhang, “Monoocc: Digging into monocular semantic occupancy prediction,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 18 398–18 405

    2024

  47. [55]

    Computing discrete fr ´echet distance,

    T. Eiter and H. Mannila, “Computing discrete fr ´echet distance,” 1994

    1994

Pith tools

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