Pith. sign in

REVIEW 4 major objections 6 minor 37 references

Pivot-Centric Trajectory Prediction: Bridging Long Horizons via Dynamical Guidance

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read PCTP claims that predicting sparse intermediate pivot points, then refining the short segments between them, reduces long-horizon motion forecasting error on Argoverse I and II across several leading models.

desk verdict PCTP is a clean, plausible decoder idea with small validation gains on multiple backbones, but the paper's leaderboard-SOTA claim is unsupported by its own tables. read the letter →

arxiv 2608.03521 v1 pith:2CXB254X submitted 2026-08-04 cs.RO cs.AI

Xiucong Zhao , Jindong Tian , Hao Miao This is my paper
classification cs.ROcs.AI
keywords trajectorypredictionmotionforecastingpivotpointslong-horizonautonomousdrivingmulti-scaledecodingArgoverseplugindecoder
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 argues that long-horizon trajectory prediction fails because decoders get either one distant endpoint to aim at or a full trajectory to refine, and proposes a middle path: predict a small set of intermediate 'pivot' points spaced uniformly in time, then predict the fine waypoints between consecutive pivots as offsets from them. The authors claim this pivot-centric decoding gives the model global intention at several moments while keeping local refinement short, so compounding errors do not build up over the full horizon. They test the idea as a plug-in decoder on four state-of-the-art models and report consistent gains on Argoverse I and II, with QCNet plus PCTP exceeding published ensemble-free results on the Argoverse II leaderboard at submission. If true, the contribution is a simple decoding change that can be dropped into existing models rather than a new end-to-end architecture.

What carries the argument

The pivot is a waypoint sampled from the ground-truth future trajectory with a time skip interval Δt; multi-scale pivots come from a hierarchy of intervals where the coarsest level is just the endpoint and finer levels add intermediate points. The pivot decoder reuses the mode-to-scene cross-attention module: at each level the predicted pivots are Fourier-embedded and fused back into the queries, so finer scales refine coarser intentions. The pivot-guided trajectory decoder predicts offsets δ between consecutive pivots and adds them to the pivot positions, reconstructing the long trajectory from short pivot-anchored segments. Gradient flow through the pivots is stopped during training, and t

What would settle it

Measure PCTP's error separately on trajectory segments that lie between pivot times in scenarios annotated with hard braking or sharp turns; if the method's gain over an endpoint-only baseline disappears or reverses on those segments, uniform temporal pivots are not providing the intermediate guidance claimed. A simpler check: double the pivot sampling interval or shift its phase by half an interval and see whether b-minFDE6 degrades sharply.

Watch

Extended reading notes

Core claim

PCTP's central claim is that the right intermediate representation for long-horizon motion forecasting is a sparse set of pivots obtained by uniformly sampling the trajectory in time at several scales. The decoder first predicts these pivot-level trajectories using global map context and agent interactions through decoupled attention; then a pivot-guided trajectory decoder predicts each fine segment as an offset from its local pivot, effectively converting one long prediction into several short ones. The paper argues that this factorization preserves the diversity of goal-based anchors and the local accuracy of refinement, while avoiding both weak endpoint guidance and whole-trajectory compo

Load-bearing premise

The method's load-bearing premise is that pivots chosen by uniformly sampling the trajectory in time capture the decisive dynamics of the motion; if a critical maneuver (sharp turn, stop, or interaction) falls between two sampled pivot times, the pivot-level guidance is blind to it and the local decoder must recover it from context alone.

Editorial extensions

If this is right

  • Any one-shot or k-shot decoder can absorb PCTP by swapping its mode-to-context attention for pivot-to-context attention, so gains transfer without retraining the backbone encoder.
  • If the reported gains hold, refinement iterations become largely unnecessary: PCTP already supplies the local guidance that iterative refinement was added for.
  • Longer horizons become easier to supervise because pivot-level matching gives the model explicit intermediate targets at multiple time scales.
  • PCTP keeps model size nearly unchanged, since it reuses the existing attention module and only changes the decoding strategy.

Reading between the lines

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

  • Pivots are defined purely by uniform temporal sampling; an obvious extension is to adapt the sampling to map topology or interaction hotspots so that a sharp turn or a yield gets its own pivot.
  • The method's success suggests a broader principle: coarse-to-fine decoding in time, not just in space, is what limits compounding error; the same idea could be tried for other long-horizon sequence prediction tasks.
  • A direct testable extension is to measure whether PCTP's advantage concentrates on scenarios with multiple sub-maneuvers (turns, lane changes, stops) and vanishes on straight-line cruising, which would confirm the mechanism it claims.
  • Because PCTP reduces refinement iterations, it may lower inference latency in real-time autonomy stacks, a consequence the paper notes only qualitatively.
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

4 major / 6 minor

Summary. The paper proposes Pivot-Centric Trajectory Prediction (PCTP), a two-stage decoder plugin for trajectory forecasting. PCTP first predicts a set of intermediate 'pivot' points at multiple temporal scales, then refines short trajectory segments between pivots using local context. The method is designed to be integrated into existing backbones by replacing their decoding heads. Experiments on Argoverse 1 and Argoverse 2 with LaFormer, HPNet, DenseTNT, and QCNet report improvements on several metrics, and the paper claims that PCTP plus QCNet outperforms all published ensemble-free methods on the Argoverse 2 test leaderboard. The central contribution is a flexible intermediate-guidance mechanism that is claimed to reduce compounding errors in long-horizon prediction.

Significance. If the empirical claims hold, PCTP would be a useful plug-in decoding strategy: it requires no new input modalities, can be dropped into multiple existing architectures, and the provided 5-run robustness check for QCNet on Argoverse 2 is a welcome step beyond single-run reporting. The method is evaluated on external benchmarks against strong baselines, and the code is released. However, the paper's strongest claim—leaderboard-level state of the art on Argoverse 2—is not supported by any in-manuscript leaderboard comparison. In addition, the paper's wording that PCTP improves accuracy 'of all considered state-of-the-art methods' is contradicted by several worsened miss-rate cells in Table 1. The claimed 'minimal impact on model size' is never measured. These are load-bearing issues because they affect the headline contributions, though the core idea remains plausible and worth further investigation.

major comments (4)
  1. [Abstract; §4.2, Table 2] The claim that 'PCTP combined with QCNet outperforms all published ensemble-free methods on the Argoverse II leaderboard at submission' is not verifiable from the manuscript. Table 2 reports only QCNet and QCNet w/ Ours; it provides no comparison to any other published method, no leaderboard snapshot, and no submission date. Because this is repeated as a key contribution in the Abstract, Introduction, and Conclusion, the claim must either be removed or substantiated with a dated leaderboard table listing all ensemble-free methods and their scores. As written, the paper's strongest empirical assertion depends entirely on external data not provided.
  2. [§4.2, Table 1] The text states that PCTP 'can improve the accuracy of all considered state-of-the-art methods' and that it 'improves most metrics,' but several miss-rate cells worsen: LaFormer MR6 degrades from 0.096 to 0.099 and MR1 from 0.474 to 0.477; HPNet MR6 degrades from 0.078 to 0.081; QCNet MR1 degrades from 0.579 to 0.584. Since the paper claims general prediction-accuracy improvement, these regressions should be acknowledged and discussed. The authors should either report per-metric consistency or qualify the claim to the metrics that improve. This is not a fatal flaw, but the selective presentation needs correction.
  3. [Abstract; §4.2] The abstract claims PCTP achieves 'minimal impact on model size,' yet no parameter counts, FLOPs, or inference-time measurements are reported anywhere in the manuscript. Given that the method adds pivot decoders and multi-scale iterations, the parameter overhead is not obvious. The authors should provide a table comparing model size and, ideally, latency before making this claim.
  4. [§3.3] The description of the one-shot variant is ambiguous and potentially inconsistent with the multi-scale pivot mechanism. The text says that for one-shot decoders 'PCTP only learns top-level pivots without iterative operations.' If 'top-level' means the coarsest level (i.e., the endpoint), then the one-shot variant reduces to a single endpoint guide and does not provide intermediate pivot guidance. If it means all pivot levels, the phrase should be clarified. This matters because Table 1 reports gains on DenseTNT, which is an endpoint-based one-shot method, and the reader cannot tell whether the claimed intermediate-guidance mechanism is actually used there.
minor comments (6)
  1. [§4.3] The text contains a broken reference 'Table reftable3' and inconsistent table numbering: the robustness experiment is described as 'Table 4' but the corresponding table is labeled Table 3. Please fix cross-references and table numbering.
  2. [§1, Contributions] The first contribution calls the method 'Pivot-Centric Trajectory Evolution (PCTP)' while the title and abstract use 'Prediction.' Please make the terminology consistent.
  3. [§1, second paragraph] The phrase 'cause map information' should likely be 'coarse map information' or 'map information.'
  4. [§4.2] The sentence 'We also present some qualitative results on the Argoverse II test set' describes Table 2, which is quantitative. Change 'qualitative' to 'quantitative.'
  5. [§4.4] The case illustration describes six modalities, but Figure 4 shows six colored trajectories; consider labeling the figure more explicitly so the reader can map the text description to the visualization.
  6. [References] Several references to 2026 venues (e.g., Tian et al. 2026, Liu et al. 2026, Xu et al. 2026) may be difficult for readers to verify. Please confirm these are published or accepted, or cite publicly available versions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: PCTP's pivots are supervised intermediate targets and the final trajectory is a learned residual around predicted pivots, tested on external benchmarks.

full rationale

The claimed derivation chain is not circular. Pivots are defined by temporal subsampling of the ground-truth trajectory (Eq. 4) and are learned as intermediate targets (Eqs. 6-10, 16). The final trajectory is produced by adding learned offsets to predicted pivots (Eqs. 12-13), and both pivot-level and point-level losses are supervised by ground truth (Eqs. 14-17). No equation forces the prediction to equal a fitted input by construction; the gradient stop through pivots is a training design, not a circular reduction. The method is evaluated against external Argoverse 1/2 benchmarks and integrated into external backbones (LaFormer, HPNet, DenseTNT, QCNet), so the empirical claims are self-contained. The paper contains several self-citations in the introduction and conclusion, but they are background references and do not provide load-bearing support for PCTP. The headline claim about outperforming all published ensemble-free methods on the Argoverse II leaderboard is not substantiated by Table 2, which only compares QCNet variants; however, that is an evidentiary gap, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central contribution is an architecture, not a derived law, so the ledger is dominated by design choices and domain assumptions about pivot semantics. The pivot sampling schedule, loss balancing, and gradient-stopping rule are hand-selected, and the load-bearing assumption that temporal subsampling captures all important maneuver information is not independently tested.

free parameters (4)
  • Number of pivot levels L = unspecified in main text; figures show 3 levels
    The hierarchical pivot schedule in Eq. 5 depends on L; the main text does not give the value used in experiments.
  • Interval growth factors alpha_l = unspecified
    Eq. 5 defines the skip-interval recursion but no numeric alpha_l values are provided in the main text.
  • Classification loss weight beta = unspecified
    Eq. 17 combines pivot loss, trajectory loss, and classification loss with a balancing hyperparameter beta that is not reported.
  • Number of modes K = 6 in experiments
    The benchmark uses K=6 trajectories; the choice of query count affects capacity and diversity but is not justified by an analysis.
assumptions (4)
  • domain assumption Uniformly time-subsampled ground-truth waypoints define a sufficient set of pivot points.
    Section 3.2 and Eq. 4 define pivots as temporal samples and state that "pivots focusing solely on temporal dynamics are sufficient for most scenarios," without quantifying failure cases.
  • domain assumption A cross-attention decoder with K learnable queries can regress accurate multi-scale pivots from agent and map features.
    Eqs. 6-10 reuse DETR-style decoupled attention and multi-scale iteration; this is inherited from prior literature and assumed to transfer to pivot regression.
  • domain assumption Stopping the gradient flow through the pivots during trajectory decoding stabilizes training without harming accuracy.
    Section 3.4 states "For stabilization, the pivot-guided trajectory decoding module stops the gradient flow through the pivots," but no ablation compares this choice to allowing gradients through.
  • domain assumption Winner-takes-all matching at the pivot level enforces intention consistency between pivot and point trajectories.
    Eq. 15 applies WTA matching to pivots; the assumption that pivot-level matching transfers to final-trajectory diversity is taken from Lee et al. without task-specific proof.
invented entities (1)
  • pivot points
    purpose: Intermediate temporally-sampled anchors that split long-horizon prediction into coarse pivot prediction and fine-grained pivot-guided offset regression.
    Pivots are defined operationally in Section 3.2 as uniformly sampled ground-truth waypoints. They have no independent falsifiable handle beyond the benchmark performance of the final network, and the paper itself asserts they contain all necessary trajectory information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pivot-Centric Trajectory Prediction: Bridging Long Horizons via Dynamical Guidance." pith.science (2026). https://pith.science/paper/2CXB254X

@misc{pith2026260803521,
  author       = {Pith},
  title        = {Pith review of: Pivot-Centric Trajectory Prediction: Bridging Long Horizons via Dynamical Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2CXB254X}},
  note         = {Machine review of arXiv:2608.03521}
}
read the original abstract

Forecasting precise future motion of surrounding agents is essential for reliable autonomous vehicles. However, as the demand for longer prediction horizons increases, existing endpoint-completion or iterative-refine methods increasingly struggle with weak guidance and compounding errors. To tackle the long-horizon prediction challenge, we propose Pivot-Centric Trajectory Prediction (PCTP). By introducing ``pivots'' and focusing on predicting pivot points along extended trajectories, we divide the long-term prediction task into short-term sub-tasks at various scales. Specifically, PCTP decouples the long-term trajectory predicting process into two processes: pivot prediction and pivot-based trajectory refinement. The pivot prediction process aims to utilize global map context and agent-to-agent interactions to identify these ``pivot points'', while the pivot-based trajectory refinement process focuses on local map details and refines the short-term trajectory based on predicted ``pivot points''. Compared with existing methods, PCTP provides more intermediate guidance while reducing compounding errors. Moreover, PCTP is a flexible approach that can be integrated into most state-of-the-art trajectory prediction models. Experimental results show that PCTP improves the prediction accuracy of leading models on both Argoverse I and Argoverse II datasets with minimal impact on model size. Specifically, PCTP combined with QCNet outperforms all published ensemble-free methods on the Argoverse II leaderboard at submission.

Figures

Figures reproduced from arXiv: 2608.03521 by the authors.

Figure 1
Figure 1. Illustration of various trajectory prediction methods: (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of PCTP. The top section illustrates the overall pipeline, while the lower section details our Multi-Scale Pivot Prediction [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Sampling method for multi-scale pivots. The three point [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of PCTP learning process. Orange polyline denotes the agent history, while the other 6 different colors represent 6 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 35 canonical work pages

  1. [1]

    Adapt: Efficient multi-agent trajec- tory prediction with adaptation

    [Aydemiret al., 2023 ] G¨orkay Aydemir, Adil Kaan Akan, and Fatma G ¨uney. Adapt: Efficient multi-agent trajec- tory prediction with adaptation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295–8305,

    work page 2023

  2. [5]

    Path-aware graph attention for hd maps in motion prediction

    [Da and Zhang, 2022] Fang Da and Yu Zhang. Path-aware graph attention for hd maps in motion prediction. In 2022 International Conference on Robotics and Automa- tion (ICRA), pages 6430–6436. IEEE,

    work page 2022

  3. [7]

    Home: Heatmap output for future motion estimation

    [Gilleset al., 2021 ] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Home: Heatmap output for future motion estimation. In2021 IEEE International Intelligent Trans- portation Systems Conference (ITSC), pages 500–507. IEEE,

    work page 2021

  4. [9]

    Densetnt: End-to-end trajectory prediction from dense goal sets

    [Guet al., 2021 ] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15303–15312,

    work page 2021

  5. [10]

    Rethinking integration of prediction and planning in deep learning- based automated driving systems: a review.arXiv preprint,

    [Hagedornet al., 2023 ] Steffen Hagedorn, Marcel Hallgar- ten, Martin Stoll, and Alexandru Condurache. Rethinking integration of prediction and planning in deep learning- based automated driving systems: a review.arXiv preprint,

    work page 2023

  6. [11]

    Planning-oriented autonomous driving

    [Huet al., 2023 ] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17853–17862,

    work page 2023

  7. [12]

    Motiondiffuser: Controllable multi-agent mo- tion prediction using diffusion

    [Jianget al., 2023 ] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent mo- tion prediction using diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9644–9653,

    work page 2023

  8. [13]

    Stochastic multiple choice learning for training diverse deep ensembles.Advances in Neural Information Processing Systems, 29,

    [Leeet al., 2016 ] Stefan Lee, Senthil Purushwalkam Shiva Prakash, Michael Cogswell, Viresh Ranjan, David Crandall, and Dhruv Batra. Stochastic multiple choice learning for training diverse deep ensembles.Advances in Neural Information Processing Systems, 29,

    work page 2016

Show all 37 references
  1. [15]

    Bat: Behavior-aware human-like trajec- tory prediction for autonomous driving

    [Liaoet al., 2024 ] Haicheng Liao, Zhenning Li, Huanming Shen, Wenxuan Zeng, Dongping Liao, Guofa Li, and Chengzhong Xu. Bat: Behavior-aware human-like trajec- tory prediction for autonomous driving. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, p...

    2024

  2. [17]

    Bus-conditioned zero-shot tra- jectory generation via task arithmetic.arXiv preprint arXiv:2602.13071,

    [Liuet al., 2026 ] Shuai Liu, Ning Cao, Yile Chen, Yue Jiang, and Gao Cong. Bus-conditioned zero-shot tra- jectory generation via task arithmetic.arXiv preprint arXiv:2602.13071,

    2026

  3. [20]

    Scene transformer: A unified ar- chitecture for predicting multiple agent trajectories.arXiv preprint arXiv:2106.08417,

    [Ngiamet al., 2021 ] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified ar- chitecture for predicting multiple agent trajectories.a...

    2021 arXiv

  4. [21]

    Covernet: Multimodal behavior prediction using trajectory sets

    [Phan-Minhet al., 2020 ] Tung Phan-Minh, Elena Corina Grigore, Freddy A Boulton, Oscar Beijbom, and Eric M Wolff. Covernet: Multimodal behavior prediction using trajectory sets. InProceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition (CVPR), pages...

    2020

  5. [22]

    Trajeglish: Traffic modeling as next-token prediction

    [Philionet al., 2024 ] Jonah Philion, Xue Bin Peng, and Sanja Fidler. Trajeglish: Traffic modeling as next-token prediction. InThe Twelfth International Conference on Learning Representations (ICLR),

    2024

  6. [23]

    Trajectron++: Dynamically-feasible trajectory forecasting with heteroge- neous data

    [Salzmannet al., 2020 ] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectron++: Dynamically-feasible trajectory forecasting with heteroge- neous data. InComputer Vision–ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Pro- cee...

    2020

  7. [24]

    Motion transformer with global intention localization and local movement refinement.Advances in Neural Information Processing Systems, 35:6531–6543,

    [Shiet al., 2022 ] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement.Advances in Neural Information Processing Systems, 35:6531–6543,

    2022

  8. [25]

    Hpnet: Dynamic trajectory forecasting with historical prediction attention

    [Tanget al., 2024 ] Xiaolong Tang, Meina Kan, Shiguang Shan, Zhilong Ji, Jinfeng Bai, and Xilin Chen. Hpnet: Dynamic trajectory forecasting with historical prediction attention. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1...

    2024

  9. [26]

    Arrow: An adap- tive rollout and routing method for global weather fore- casting

    [Tianet al., 2026 ] Jindong Tian, Yifei Ding, Ronghui Xu, Hao Miao, Chenjuan Guo, and Bin Yang. Arrow: An adap- tive rollout and routing method for global weather fore- casting. InThe Fourteenth International Conference on Learning Representations (ICLR),

    2026

  10. [27]

    Identifying driver interactions via conditional behavior prediction

    [Tolstayaet al., 2021 ] Ekaterina Tolstaya, Reza Mahjourian, Carlton Downey, Balakrishnan Vadarajan, Benjamin Sapp, and Dragomir Anguelov. Identifying driver interactions via conditional behavior prediction. In2021 IEEE Inter- national Conference on Robotics and Automation (IC...

    2021

  11. [28]

    Multi- path++: Efficient information fusion and trajectory aggre- gation for behavior prediction

    [Varadarajanet al., 2022 ] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douil- lard, Chi Pang Lam, Dragomir Anguelov, et al. Multi- path++: Efficient information fusion and trajectory aggre- ga...

    2022

  12. [29]

    C2f- tp: A coarse-to-fine denoising framework for uncertainty- aware trajectory prediction

    [Wanget al., 2025 ] Zichen Wang, Hao Miao, Senzhang Wang, Renzhi Wang, Jianxin Wang, and Jian Zhang. C2f- tp: A coarse-to-fine denoising framework for uncertainty- aware trajectory prediction. InProceedings of the AAAI conference on artificial intelligence, volume 39, pages 12...

    2025

  13. [30]

    Most: A foundation model for multi-modality spatio-temporal traffic prediction

    [Xuet al., 2026 ] Ronghui Xu, Jihao Chen, Jindong Tian, Chenjuan Guo, and Bin Yang. Most: A foundation model for multi-modality spatio-temporal traffic prediction. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 1 (KDD), pages 1716–1727,

    2026

  14. [31]

    Dcms: Mo- tion forecasting with dual consistency and multi-pseudo- target supervision.arXiv preprint arXiv:2204.05859,

    [Yeet al., 2022 ] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Dcms: Mo- tion forecasting with dual consistency and multi-pseudo- target supervision.arXiv preprint arXiv:2204.05859,

    2022 arXiv

  15. [32]

    Bootstrap motion forecasting with self-consistent constraints

    [Yeet al., 2023 ] Maosheng Ye, Jiamiao Xu, Xunnong Xu, Tengfei Wang, Tongyi Cao, and Qifeng Chen. Bootstrap motion forecasting with self-consistent constraints. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision, pages 8504–8514,

    2023

  16. [33]

    Lanercnn: Distributed repre- sentations for graph-centric motion forecasting

    [Zenget al., 2021 ] Wenyuan Zeng, Ming Liang, Renjie Liao, and Raquel Urtasun. Lanercnn: Distributed repre- sentations for graph-centric motion forecasting. In2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 532–539. IEEE,

    2021

  17. [34]

    Tnt: Target- driven trajectory prediction

    [Zhaoet al., 2021 ] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Ben Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yuning Chai, Cordelia Schmid, et al. Tnt: Target- driven trajectory prediction. InConference on Robot Learning, pages 895–904. PMLR,

    2021

  18. [35]

    Hivt: Hierarchical vector trans- former for multi-agent motion prediction

    [Zhouet al., 2022 ] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector trans- former for multi-agent motion prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 8823–8833,

    2022

  19. [36]

    Query-centric trajectory predic- tion

    [Zhouet al., 2023 ] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory predic- tion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17863–17873,

    2023

  20. [37]

    Smartre- fine: A scenario-adaptive refinement framework for effi- cient motion prediction

    [Zhouet al., 2024 ] Yang Zhou, Hao Shao, Letian Wang, Steven L Waslander, Hongsheng Li, and Yu Liu. Smartre- fine: A scenario-adaptive refinement framework for effi- cient motion prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (C...

    2024

  21. [2016]

    Learning lane graph representations for motion forecast- ing

    [Lianget al., 2020 ] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecast- ing. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceed- ings, Part II...

    2020

  22. [2019]

    Gorela: Go rel- ative for viewpoint-invariant motion forecasting

    [Cuiet al., 2023 ] Alexander Cui, Sergio Casas, Kelvin Wong, Simon Suo, and Raquel Urtasun. Gorela: Go rel- ative for viewpoint-invariant motion forecasting. In2023 IEEE International Conference on Robotics and Automa- tion (ICRA), pages 7801–7807. IEEE,

    2023

  23. [2020]

    Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction.arXiv preprint arXiv:1910.05449,

    [Chaiet al., 2019 ] Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction.arXiv preprint arXiv:1910.05449,

    2019 arXiv

  24. [2021]

    Gohome: Graph-oriented heatmap output for future motion estimation

    [Gilleset al., 2022 ] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Gohome: Graph-oriented heatmap output for future motion estimation. In2022 international conference on robotics and automation (ICRA), pages 9107–9114. IEEE,

    2022

  25. [2022]

    Vectornet: Encoding hd maps and agent dy- namics from vectorized representation

    [Gaoet al., 2020 ] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dy- namics from vectorized representation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

    2020

  26. [2023]

    End-to-end object detection with transformers

    [Carionet al., 2020 ] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean conference on computer vision (ECCV), pages 213–229. Springer,

    2020

  27. [2024]

    Laformer: Trajectory pre- diction for autonomous driving with lane-aware scene con- straints

    [Liuet al., 2024 ] Mengmeng Liu, Hao Cheng, Lin Chen, Hellward Broszio, Jiangtao Li, Runjiang Zhao, Monika Sester, and Michael Ying Yang. Laformer: Trajectory pre- diction for autonomous driving with lane-aware scene con- straints. InProceedings of the IEEE/CVF Conference on C...

    2024

  28. [2025]

    Wayformer: Motion forecasting via simple & efficient attention networks

    [Nayakantiet al., 2023 ] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Ben- jamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 29...

    2023

  29. [2026]

    Find3: A dual 3d state space model with dynamic hypergraph for fi- nancial stock prediction

    [Meiet al., 2025 ] Jieyuan Mei, Jindong Tian, Ronghui Xu, Hanyue Wei, Chenjuan Guo, and Bin Yang. Find3: A dual 3d state space model with dynamic hypergraph for fi- nancial stock prediction. InProceedings of the 34th ACM International Conference on Information and Knowledge Ma...

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.