Pith. sign in

REVIEW 3 major objections 5 minor 32 references

LTMSformer: A Local Trend-Aware Attention and Motion State Encoding Transformer for Multi-Agent Trajectory Prediction

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LTMSformer claims that explicitly modeling local temporal trends and high-order motion states—acceleration, jerk, heading—improves multi-agent trajectory prediction, reporting lower minADE, minFDE, and miss rate than the HiVT-64 baseline…

desk verdict LTMSformer is a competent, incremental HiVT improvement where the MLP refinement and motion-state modules look real, but the headline local-trend attention module is underspecified and its ablation gain is inside single-seed noise. read the letter →

arxiv 2507.04634 v1 pith:JXF5HDTT submitted 2025-07-07 cs.CV cs.AI

Yixin Yan , Yang Li , Yuanfan Wang , Xiaozhou Zhou , Beihao Xia , Manjiang Hu , Hongmao Qin This is my paper
classification cs.CVcs.AI
keywords multi-agenttrajectorypredictionlocaltrend-awareattentionmotionstateencodingArgoverse1transformerrefinementautonomousdriving
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 claims that trajectory prediction for autonomous driving improves when a model explicitly captures short-term motion trends in the observed sequence and when spatial interaction features include higher-order motion attributes such as acceleration and jerk. The authors build LTMSformer on the HiVT-64 baseline by adding three modules: a Local Trend-Aware Attention that applies causal convolutional attention inside non-overlapping local time boxes, a Motion State Encoder that feeds acceleration, jerk, and heading into agent-agent attention, and a Lightweight Proposal Refinement Module that refines six candidate trajectories with MLPs. On the Argoverse 1 benchmark they report that the full model lowers minADE by about 4.35%, minFDE by 8.74%, and miss rate by 20% relative to HiVT-64, while reaching the accuracy of the larger HiVT-128 with 68% fewer parameters. The broader point is that inexpensive inductive biases about local temporal dynamics and motion states can close part of the gap to much larger transformer models.

What carries the argument

Three components carry the argument. Local Trend-Aware Attention (LTAA) splits the observed temporal sequence into hierarchical non-overlapping local boxes of sizes 3, 7, and 21, applies causal 1×k convolutions to produce queries and keys, and runs attention within each box so that each time step attends mainly to its immediate neighbours while stacked layers widen the receptive field. The Motion State Encoder (MSE) augments each agent-pair feature with the neighbour's acceleration $a_j$, jerk $\dot{a}_j$, and heading $\theta_j$ via an MLP, then uses these as keys and values in multi-head attention with the LTAA output as query. The Lightweight Proposal Refinement Module (LPRM) embeds the stage-one trajectories and the full observed-plus-predicted sequence with MLPs, concatenates them with local and global interaction embeddings, and predicts a residual offset to refine each mode.

What would settle it

Re-run the ablation on Argoverse 1 with the LTAA module only (no MSE or LPRM), sweeping box sizes and kernel size k; if a uniform box size or a different schedule yields the same minADE as the chosen {3,7,21}, the hierarchical schedule is not load-bearing. Alternatively, train the same model on Argoverse 2's six-second horizon without re-tuning the box sizes; if the gains over HiVT vanish, the trend-aware attention is tuned to the three-second horizon rather than being a general inductive bias.

Watch

Extended reading notes

Core claim

On its own terms, LTMSformer establishes that three targeted additions to the Hierarchical Vector Transformer (HiVT) recipe—local box convolutional attention for temporal trends, motion-state-augmented spatial attention, and MLP-based single-pass refinement—jointly improve multi-modal trajectory prediction accuracy on Argoverse 1 while shrinking the model. The validation-set gains over HiVT-64 are reported as minADE 0.66 vs 0.69, minFDE 0.94 vs 1.03, and MR 0.08 vs 0.10, with the test set showing minADE 0.7993 vs 0.8306, minFDE 1.1932 vs 1.3053, and MR 0.1254 vs 0.1503. The same model matches HiVT-128's test minADE (0.7993) with 789k parameters against 2529k. Ablations attribute the improvements to each component incrementally, with the largest single jump coming from the Motion State Encoder.

Load-bearing premise

The hand-picked schedule of local time-box sizes (3, 7, 21) and the causal 1×k convolution within each box, with k unspecified, are chosen for the Argoverse 1 three-second horizon; if a different box schedule or kernel size performs equally well, the specific local-trend inductive bias is not what drives the gains.

Editorial extensions

If this is right

  • If correct, adding local temporal attention and motion-state features to a vector transformer improves all three standard metrics (minADE, minFDE, MR) over the base model.
  • The MLP-only refinement module shows that multi-stage trajectory refinement does not require GRUs or attention layers, suggesting a parameter-efficient refinement recipe.
  • The reported parameter count (789k) means the full model is about one-third the size of HiVT-128 while matching its test minADE and improving minFDE and MR.
  • On the Argoverse 1 test set, LTMSformer achieves the lowest minFDE and MR among the listed single-model methods, without ensembling or data augmentation.

Reading between the lines

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

  • The ablation table shows only a small gain from adding LTAA on top of MSE (minADE 0.673 to 0.672, minFDE 1.001 to 1.000), so the headline improvement may be carried mostly by MSE and LPRM; a clean test would be to ablate LTAA alone.
  • The box schedule {3,7,21} is tailored to Argoverse 1's 30-step observation window; extending to Argoverse 2's longer horizon would likely require re-tuning, which the paper lists as future work.
  • The claimed 'higher accuracy than HiVT-128' rests on matching its test minADE exactly while improving minFDE and MR; without repeated-run variance or significance tests, the margin should be read as indicative rather than guaranteed.
  • Since the kernel size $k$ of the causal convolution is not reported, the local-trend module is not fully specified for reproduction; reporting $k$ would make the inductive-bias claim testable.
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 / 5 minor

Summary. The paper presents LTMSformer, a multi-agent trajectory prediction model built on the HiVT architecture, augmented with three modules: Local Trend-Aware Attention (LTAA), a Motion State Encoder (MSE), and a Lightweight Proposal Refinement Module (LPRM). LTAA uses causal convolutional attention within hierarchical non-overlapping time boxes to capture local temporal trends; MSE injects acceleration, jerk, and heading into spatial interaction; and LPRM refines stage-one proposals with MLPs. On Argoverse 1, the authors report that LTMSformer reduces minADE, minFDE, and MR relative to HiVT-64 and achieves higher accuracy than HiVT-128 with 68% fewer parameters (789k vs 2529k). The validation-set ablation shows monotonic gains when adding MSE, LTAA, and LPRM, and a sensitivity study sweeps the stage-two loss weight lambda_1.

Significance. If the reported numbers are reproducible, the paper's main practical contribution is a parameter-efficient improvement over HiVT-128, with a modest model of 789k parameters. The use of official HiVT checkpoints for baselines and the inclusion of a sensitivity study are strengths. However, the paper's first claimed contribution, the LTAA module, is currently not established: its convolution kernel size is unspecified and its marginal ablation gain is within single-seed noise. The MSE module is also insufficiently defined to be reproduced. The central claim that all three modules contribute to the reported gains therefore needs additional evidence before the results can be fully credited.

major comments (3)
  1. [III.C.2, Eq. (5), Table II] The causal 1xk convolution in Eq. (5) has an unspecified kernel size k, and the local box sizes B={3,7,21} are selected ad hoc for the 20-step Argoverse 1 horizon without a sweep. This makes the LTAA module non-reproducible and prevents an assessment of whether the chosen receptive field is the right inductive bias. More importantly, the ablation in Table II attributes the smallest and least convincing gain to LTAA: adding LTAA to A1 changes minADE from 0.673 to 0.672, minFDE from 1.001 to 1.000, and MR from 0.098 to 0.096. With a single run, these differences are indistinguishable from training noise. Please specify k, report the sensitivity of the final results to k and to the box schedule, and provide multi-seed or alternative ablation evidence that LTAA contributes beyond noise.
  2. [III.C.3, Eq. (7)] The motion state attributes a_j, a_dot_j, and theta_j in Eq. (7) are not defined. It is unclear whether acceleration and jerk are computed by finite differences of observed positions, at which time index they are evaluated, and whether they are sequences or scalar values. This makes the MSE module irreproducible and makes it impossible to judge whether the comparison with HiVT-64, which encodes only relative positions, is a fair ablation of the proposed motion-state information. Please provide exact formulas, time indices, and normalization details for all entries of k_ij.
  3. [Tables I-III] All quantitative results are single runs with no error bars, confidence intervals, or significance tests. This is a particular concern for the headline comparisons: the validation-set differences between LTMSformer and HiVT-64 in Table I are 0.03 in minADE and 0.09 in minFDE, while the incremental LTAA gain in Table II is 0.001. Without multiple seeds or a paired evaluation, the reported gains could be explained by training variance. Please report mean and standard deviation over at least three independent training runs for the main results and ablations, and clearly state the metric-evaluation protocol.
minor comments (5)
  1. [Figure 3] The caption of Figure 3 says local boxes of sizes 3, 6, and 12, while the text and the implementation details use {3, 7, 21}; please correct the inconsistency.
  2. [Eq. (5)] Equation (5) writes the same BatchNorm(conv(zi)) for both Q and K; it should be clarified whether these are two independent convolutional transformations with separate weights and what the output dimensions are.
  3. [References] Reference [30] attributes the BERT paper to J. Lee and K. Toutanova, but the cited title, 'Pre-training of deep bidirectional transformers for language understanding,' is by Devlin et al.; please fix the citation.
  4. [I and III.F] The phrase 'with fewer parameters' in the description of the LPRM contribution is not quantified against a specific alternative; a per-module parameter count or a comparison with an attention-based refinement baseline would make the efficiency claim precise.
  5. [IV.A] The paper does not mention whether code or trained models will be released; providing a public implementation would mitigate the specification issues noted above.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical architecture study evaluated on a standard benchmark with external baselines; no fitted constant is disguised as a prediction.

full rationale

The paper's central claim is that LTMSformer improves trajectory prediction accuracy relative to HiVT-64 and HiVT-128 on Argoverse 1. This claim is established by training on the public Argoverse 1 dataset and comparing against HiVT checkpoints obtained from the official release, which is external evidence rather than a circular construction. The losses (cross-entropy, negative log-likelihood, Smooth L1) and metrics (minADE, minFDE, MR) are standard and independent of the proposed modules. The ablation study incrementally adds MSE, LTAA, and LPRM; although the gains attributed to LTAA are small and no repeated-seed statistics are reported, this is a legitimate empirical comparison, not a case where a fitted parameter is renamed as a prediction. The underspecified convolution kernel size k and the hand-selected local box sizes {3,7,21} are reproducibility and robustness concerns, but they do not make the derivation circular. The paper does cite prior work by overlapping authors [24], but that citation is used only as related work and is not load-bearing for any claim. No uniqueness theorem is invoked, and no central premise reduces to a self-citation. Therefore, no significant circularity is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The model inherits most of its structure from HiVT and introduces three new modules with several hand-chosen hyperparameters. There are no invented physical entities. The main free parameters are the local box schedule, the number of layers, the loss weight, and the unspecified convolution kernel size. The key domain assumptions concern the reliability of acceleration and jerk features and the comparability of baseline numbers.

free parameters (6)
  • Local box sizes B = 3, 7, 21
    Hand-chosen to fit the observed sequence length in Argoverse 1 (20 observed steps). No sensitivity study is reported for these values.
  • Causal convolution kernel size k = Not stated
    Appears in Eq. (5) as a 1xk convolution, but the value of k is never given in the paper. This is required to reproduce the LTAA module.
  • Loss weight lambda1 = 5
    Selected as the best value in the sensitivity study (Table III) on the validation set, so it is tuned to the data.
  • Number of LTAA layers = 3
    Architecture choice; the number of layers is not swept or justified separately from the box sizes.
  • Hidden units = 64
    Inherited from the HiVT-64 baseline; not varied in this paper.
  • Local region radius = 50 m
    Inherited from HiVT; not varied in this paper.
assumptions (5)
  • domain assumption Causal 1xk convolution preserves temporal causality
    Used in Eq. (5). The model assumes future time steps do not influence current features, which is standard for causal prediction but not explicitly justified.
  • ad hoc to paper Non-overlapping local boxes are a valid segmentation of temporal dependencies
    The paper introduces box sizes {3, 7, 21} without a derivation or ablation, and assumes that cross-box dependencies are not critical for prediction.
  • domain assumption Acceleration and jerk can be reliably estimated from observed positions
    Eq. (7) uses a_j and dot a_j as input features. No smoothing or noise model is described, so the assumption is that differentiating Argoverse position data produces helpful signal.
  • domain assumption Argoverse 1 benchmark and its train/val/test split are a valid evaluation protocol
    The dataset is used without citing the Argoverse paper, and the paper assumes the standard split and metrics are the right way to measure trajectory prediction.
  • domain assumption Baseline results quoted from other papers are comparable to locally run HiVT results
    Table I mixes locally reproduced HiVT-64/128 checkpoints with numbers taken from other papers. The paper does not discuss whether those baselines used identical training, validation splits, or evaluation code.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LTMSformer: A Local Trend-Aware Attention and Motion State Encoding Transformer for Multi-Agent Trajectory Prediction." pith.science (2026). https://pith.science/paper/JXF5HDTT

@misc{pith2026250704634,
  author       = {Pith},
  title        = {Pith review of: LTMSformer: A Local Trend-Aware Attention and Motion State Encoding Transformer for Multi-Agent Trajectory Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JXF5HDTT}},
  note         = {Machine review of arXiv:2507.04634}
}
read the original abstract

It has been challenging to model the complex temporal-spatial dependencies between agents for trajectory prediction. As each state of an agent is closely related to the states of adjacent time steps, capturing the local temporal dependency is beneficial for prediction, while most studies often overlook it. Besides, learning the high-order motion state attributes is expected to enhance spatial interaction modeling, but it is rarely seen in previous works. To address this, we propose a lightweight framework, LTMSformer, to extract temporal-spatial interaction features for multi-modal trajectory prediction. Specifically, we introduce a Local Trend-Aware Attention mechanism to capture the local temporal dependency by leveraging a convolutional attention mechanism with hierarchical local time boxes. Next, to model the spatial interaction dependency, we build a Motion State Encoder to incorporate high-order motion state attributes, such as acceleration, jerk, heading, etc. To further refine the trajectory prediction, we propose a Lightweight Proposal Refinement Module that leverages Multi-Layer Perceptrons for trajectory embedding and generates the refined trajectories with fewer model parameters. Experiment results on the Argoverse 1 dataset demonstrate that our method outperforms the baseline HiVT-64, reducing the minADE by approximately 4.35%, the minFDE by 8.74%, and the MR by 20%. We also achieve higher accuracy than HiVT-128 with a 68% reduction in model size.

Figures

Figures reproduced from arXiv: 2507.04634 by the authors.

Figure 1
Figure 1. Predicted trajectories are generated from the observed trajectories [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Based on HiVT [21], our method introduces three additional components, including Local Trend-Aware Attention Mechanism (LTAA), Motion [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of Local Box Convolutional Attention with local boxes [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of the ablation study result. It is carried out by incrementally adding MSE, LTAA, and LPRM, highlighting the contribution of each [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization results of LTMSformer and HiVT-64 under weak, moderate, and strong interaction scenarios. LTMSformer consistently achieves [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 30 canonical work pages

  1. [1]

    Scene transformer: A unified architecture for predicting multiple agent tra- jectories,

    J. Ngiam, B. Caine, V . Vasudevan, Z. Zhang, H.-T. L. Chiang, J. Ling, R. Roelofs, A. Bewley, C. Liu, A. Venugopal et al. , “Scene transformer: A unified architecture for predicting multiple agent tra- jectories,” 2021, arXiv:2106.08417

    arXiv 2021

  2. [2]

    Pedestrian trajectory predic- tion combining probabilistic reasoning and sequence learning,

    Y . Li, X.-Y . Lu, J. Wang, and K. Li, “Pedestrian trajectory predic- tion combining probabilistic reasoning and sequence learning,” IEEE Transactions on Intelligent Vehicles, vol. 5, no. 3, pp. 461–474, 2020

    work page 2020

  3. [3]

    A survey on trajectory-prediction methods for autonomous driving,

    Y . Huang, J. Du, Z. Yang, Z. Zhou, L. Zhang, and H. Chen, “A survey on trajectory-prediction methods for autonomous driving,” IEEE trans. intell. veh., vol. 7, no. 3, pp. 652–674, Sep. 2022

    work page 2022

  4. [4]

    Macformer: Map-agent coupled transformer for real-time and robust trajectory prediction,

    C. Feng, H. Zhou, H. Lin, Z. Zhang, Z. Xu, C. Zhang, B. Zhou, and S. Shen, “Macformer: Map-agent coupled transformer for real-time and robust trajectory prediction,” IEEE Robot. Autom. Lett. , vol. 8, no. 10, p. 6795–6802, Oct. 2023

    work page 2023

  5. [5]

    Where will the oncoming vehicle be the next second?

    A. Barth and U. Franke, “Where will the oncoming vehicle be the next second?” in Proc. IEEE Intell. Veh. Symp. (IV) , Jun. 2008

    work page 2008

  6. [6]

    Recognition of dangerous situations within a cooperative group of vehicles,

    T. Batz, K. Watson, and J. Beyerer, “Recognition of dangerous situations within a cooperative group of vehicles,” inProc. IEEE Intell. Veh. Symp. (IV), Jun. 2009, pp. 907–912

    work page 2009

  7. [7]

    Model-based threat assessment for avoiding arbitrary vehicle collisions,

    M. Br ¨annstr¨om, E. Coelingh, and J. Sj ¨oberg, “Model-based threat assessment for avoiding arbitrary vehicle collisions,” IEEE trans Intell Transp Syst., vol. 11, no. 3, pp. 658–669, Sep. 2010

    work page 2010

  8. [8]

    An integrated approach to maneuver-based trajectory prediction and criticality assessment in arbitrary road environments,

    M. Schreier, V . Willert, and J. Adamy, “An integrated approach to maneuver-based trajectory prediction and criticality assessment in arbitrary road environments,” IEEE trans Intell Transp Syst. , vol. 17, no. 10, pp. 2751–2766, Oct. 2016

    work page 2016

Show all 32 references
  1. [9]

    A game-theoretic approach to replanning-aware interactive scene prediction and planning,

    M. Bahram, A. Lawitzky, J. Friedrichs, M. Aeberhard, and D. Woll- herr, “A game-theoretic approach to replanning-aware interactive scene prediction and planning,” IEEE Trans. Veh. Technol, vol. 65, no. 6, pp. 3981–3992, Jun. 2015

    2015

  2. [10]

    Probabilistic intention prediction and trajectory generation based on dynamic bayesian net- works,

    G. He, X. Li, Y . Lv, B. Gao, and H. Chen, “Probabilistic intention prediction and trajectory generation based on dynamic bayesian net- works,” in Proc. Chin. Autom. Congr. (CAC) , Nov. 2019, pp. 2646– 2651

    2019

  3. [11]

    A dynamic bayesian network for vehicle maneuver prediction in highway driving scenarios: Framework and verification,

    J. Li, B. Dai, X. Li, X. Xu, and D. Liu, “A dynamic bayesian network for vehicle maneuver prediction in highway driving scenarios: Framework and verification,” Electronics, vol. 8, no. 1, p. 40, Jan. 2019

    2019

  4. [12]

    Social lstm: Human trajectory prediction in crowded spaces,

    A. Alahi, K. Goel, V . Ramanathan, A. Robicquet, L. Fei-Fei, and S. Savarese, “Social lstm: Human trajectory prediction in crowded spaces,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , Jun. 2016, pp. 961–971

    2016

  5. [13]

    Traphic: Tra- jectory prediction in dense and heterogeneous traffic using weighted interactions,

    R. Chandra, U. Bhattacharya, A. Bera, and D. Manocha, “Traphic: Tra- jectory prediction in dense and heterogeneous traffic using weighted interactions,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2019, pp. 8483–8492

    2019

  6. [14]

    Trafficpredict: Trajectory prediction for heterogeneous traffic-agents,

    Y . Ma, X. Zhu, S. Zhang, R. Yang, W. Wang, and D. Manocha, “Trafficpredict: Trajectory prediction for heterogeneous traffic-agents,” in Proc. AAAI Conf. Artif. Intell. (AAAI) , vol. 33, no. 01, Jul. 2019, pp. 6120–6127

    2019

  7. [15]

    Vectornet: Encoding hd maps and agent dynamics from vectorized representation,

    J. Gao, C. Sun, H. Zhao, Y . Shen, D. Anguelov, C. Li, and C. Schmid, “Vectornet: Encoding hd maps and agent dynamics from vectorized representation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2020, pp. 11 525–11 533

    2020

  8. [16]

    Leveraging future relationship reasoning for vehicle trajectory prediction,

    D. Park, H. Ryu, Y . Yang, J. Cho, J. Kim, and K.-J. Yoon, “Leveraging future relationship reasoning for vehicle trajectory prediction,” Proc. Int. Conf. Learn. Represent. , 2023

    2023

  9. [17]

    Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning,

    C. Xu, M. Li, Z. Ni, Y . Zhang, and S. Chen, “Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2022, pp. 6498–6507

    2022

  10. [18]

    Hdgt: Heterogeneous driving graph transformer for multi-agent trajectory prediction via scene encoding,

    X. Jia, P. Wu, L. Chen, Y . Liu, H. Li, and J. Yan, “Hdgt: Heterogeneous driving graph transformer for multi-agent trajectory prediction via scene encoding,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 45, no. 11, p. 13860–13875, Nov. 2023

    2023

  11. [19]

    Gsan: Graph self-attention network for learning spatial–temporal interaction representation in autonomous driving,

    L. Ye, Z. Wang, X. Chen, J. Wang, K. Wu, and K. Lu, “Gsan: Graph self-attention network for learning spatial–temporal interaction representation in autonomous driving,” IEEE Internet Things J., vol. 9, no. 12, pp. 9190–9204, Jun. 2021

    2021

  12. [20]

    A novel transformer-based model for motion forecasting in connected automated vehicles,

    M. N. Azadani and A. Boukerche, “A novel transformer-based model for motion forecasting in connected automated vehicles,” in ICC 2024- IEEE Int. Conf. Commun. , Jun. 2024, pp. 5359–5364

    2024

  13. [21]

    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,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , Jun. 2022, pp. 8823–8833

    2022

  14. [22]

    Wayformer: Motion forecasting via simple & efficient attention networks,

    N. Nayakanti, R. Al-Rfou, A. Zhou, K. Goel, K. S. Refaat, and B. Sapp, “Wayformer: Motion forecasting via simple & efficient attention networks,” in 2023 IEEE Int. Conf. Robot. Autom. (ICRA) , May 2023, pp. 2980–2987

    2023

  15. [23]

    Hierarchical vector transformer vehicle trajectories prediction with diffusion convolutional neural networks,

    Y . Tang, H. He, and Y . Wang, “Hierarchical vector transformer vehicle trajectories prediction with diffusion convolutional neural networks,” Neurocomputing, vol. 580, p. 127526, May 2024

    2024

  16. [24]

    A lightweight lane-guided vector transformer for multi-agent trajectory prediction in autonomous driving,

    Y . Pu, Y . Li, B. Xia, X. Wang, H. Qin, and L. Zhu, “A lightweight lane-guided vector transformer for multi-agent trajectory prediction in autonomous driving,” in 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC) , 2024, pp. 1245–1252

    2024

  17. [25]

    Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints,

    M. Liu, H. Cheng, L. Chen, H. Broszio, J. Li, R. Zhao, M. Sester, and M. Y . Yang, “Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2024, pp. 2039–2049

    2024

  18. [26]

    Prophnet: Efficient agent-centric motion forecasting with anchor-informed proposals,

    X. Wang, T. Su, F. Da, and X. Yang, “Prophnet: Efficient agent-centric motion forecasting with anchor-informed proposals,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , Jun. 2023, pp. 21 995– 22 003

    2023

  19. [27]

    Motion transformer with global intention localization and local movement refinement,

    S. Shi, L. Jiang, D. Dai, and B. Schiele, “Motion transformer with global intention localization and local movement refinement,” Adv. neural inf. process. syst. , vol. 35, pp. 6531–6543, 2022

    2022

  20. [28]

    Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction,

    B. Varadarajan, A. Hefny, A. Srivastava, K. S. Refaat, N. Nayakanti, A. Cornman, K. Chen, B. Douillard, C. P. Lam, D. Anguelov et al. , “Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction,” in 2022 IEEE Int. Conf. Robot. Autom. (ICRA),...

    2022

  21. [29]

    R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement,

    S. Choi, J. Kim, J. Yun, and J. W. Choi, “R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), Oct. 2023, pp. 8525–8535

    2023

  22. [30]

    Pre-training of deep bidirectional trans- formers for language understanding,

    J. Lee and K. Toutanova, “Pre-training of deep bidirectional trans- formers for language understanding,” 2018, arXiv:1810.04805

    2018 arXiv

  23. [31]

    Densetnt: End-to-end trajectory pre- diction from dense goal sets,

    J. Gu, C. Sun, and H. Zhao, “Densetnt: End-to-end trajectory pre- diction from dense goal sets,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), Oct. 2021, pp. 15 303–15 312

    2021

  24. [32]

    Ltp: Lane-based trajectory prediction for autonomous driving,

    J. Wang, T. Ye, Z. Gu, and J. Chen, “Ltp: Lane-based trajectory prediction for autonomous driving,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2022, pp. 17 134–17 142

    2022

Pith tools

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