Pith. sign in

REVIEW 3 major objections 6 minor 26 references

Goal-based Trajectory Prediction for improved Cross-Dataset Generalization

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that trajectory prediction generalizes better to unseen areas when the model chooses a goal lane, then a goal point on that lane, before completing the trajectory.

desk verdict A solid architectural contribution with a plausible mechanism, but the cross-dataset generalization claim rests on a single transfer direction and needs broader evidence before it can be taken as a general result. read the letter →

arxiv 2507.18196 v1 pith:WILBYQ6F submitted 2025-07-24 cs.LG

classification cs.LG
keywords trajectorypredictiongoal-basedcross-datasetgeneralizationgraphneuralnetworkmulti-stagegoalselectionautonomousdrivingoff-roadrate
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 a trajectory prediction model generalizes better to unseen driving environments when its endpoint is chosen in stages: first a lane on the road network, then a goal point on that lane, rather than a trajectory predicted directly. In a cross-dataset test, training on Argoverse2 and evaluating on NuScenes, the staged-goal model's best final displacement error degrades by only 8%, compared with 20.6% for Wayformer and 30.1% for MTR. If the result holds, road-bound goal selection acts as an inductive bias that keeps predictions on the road in unfamiliar areas, reducing off-road rates and making deployment in new cities safer.

What carries the argument

The multi-stage goal selection is the load-bearing mechanism. Agent-query nodes, one per predicted mode, are connected first to accessible lane-nodes through '(agent-query, decide, lane)' edges; after a lane is chosen, they are connected to the point-nodes of that lane through '(agent-query, decide, point)' edges, and the highest-scoring point becomes the goal. Non-road-bound agents use '(agent-query, decide, point-nrb)' edges to select from artificial points placed on concentric circles scaled by mean velocity. Each logit is computed as an MLP over the concatenated node and edge features plus the edge feature itself, and a softmax over all goals of a query node yields scores. This staged, graph-native scoring is what the authors credit for keeping predictions on the road and lowering the cross-dataset error drop.

What would settle it

Train HoliGraph:Goal on NuScenes and evaluate on Argoverse2, or evaluate on a held-out city within Argoverse2; if the staged-goal model does not show the smallest relative performance drop among the same baselines, the generalization claim fails. Alternatively, re-running the Argoverse2-to-NuScenes transfer with non-linear or learned interpolation of the 2 Hz NuScenes data would test whether the reported 8% drop depends on the linear upscaling procedure.

Watch

Extended reading notes

Core claim

HoliGraph:Goal is an encoder-decoder graph neural network that predicts future trajectories for road-bound and non-road-bound agents. For road-bound agents (vehicles, cyclists, trucks), the decoder first classifies a goal lane among all reachable lanes, then classifies a point-node on that lane, then regresses a local offset and completes a trajectory to the refined goal; pedestrians select an artificial point from concentric circles generated around the agent. The paper's central finding is that this multi-stage goal selection reduces the relative performance drop in cross-dataset transfer: trained on Argoverse2 and evaluated on NuScenes, the model's b-minFDE6 worsens by 8% (3.13 m to 3.38 m), minFDE6 by 10.3%, minADE6 by 5.6%, minMR6 by 5.4%, and the off-road rate stays at 0.03, whereas the compared baselines drop substantially more. The authors conclude that softly binding trajectories to the road network through staged goal selection improves generalization to unseen scenarios.

Load-bearing premise

The conclusion that staged goal selection improves generalization to unseen areas rests on a single transfer direction, training on Argoverse2 and testing on NuScenes, and assumes that this one dataset pair stands for arbitrary unseen environments.

Editorial extensions

If this is right

  • A model trained on Argoverse2 and deployed on NuScenes would keep its final displacement error increase to 8%, compared with 20-30% for Wayformer, Autobot, and MTR.
  • Off-road rate, which often jumps under distribution shift, stays at 0.03 for HoliGraph:Goal across the transfer, meaning staged goal selection prevents the typical off-road failure mode on unseen maps.
  • On the smaller NuScenes dataset, the multi-stage goal selection improves waypoint accuracy and off-road rate over the direct-prediction HoliGraph baseline, suggesting the benefit is strongest when training data is limited.
  • Because the trajectory scores are taken directly from the goal classification, a dedicated scoring module would likely further improve metrics that depend on mode confidence, as the authors note for future work.

Reading between the lines

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

  • A natural next test is the reverse transfer, NuScenes to Argoverse2; if the staged-goal model no longer shows the smallest relative drop, the claimed generalization benefit may be tied to this specific dataset pair.
  • The NuScenes evaluation relies on linear interpolation from 2 Hz to 10 Hz; varying the interpolation method would show how much of the 8% advantage depends on the upscaling scheme.
  • The same staged goal-selection idea could be applied to other domains with hard geometric output constraints, such as robot navigation on waypoint graphs or lane keeping in structured environments.
  • Since the model already gains robustness from translational and rotational invariance of relative features, combining staged goals with stronger equivariance could compound the generalization improvement.
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 HoliGraph:Goal, a heterogeneous graph neural network for multimodal trajectory prediction. Its main novelty is a multi-stage, map-bound goal selection process: road-bound agents first classify a goal lane and then a goal point on that lane, while non-road-bound agents select from artificially generated point-nrb nodes; the selected goal is locally refined and used to complete trajectories. The authors claim that this staged goal selection improves cross-dataset generalization. They evaluate on Argoverse2 and NuScenes, and report that, when training on Argoverse2 and evaluating on NuScenes, HoliGraph:Goal has smaller relative performance drops than the HoliGraph baseline and than the SotA models reported in UniTraj (Table III). Same-dataset results on Argoverse2 (Table I) and NuScenes (Table II) are also provided, along with qualitative examples showing fewer off-road trajectories.

Significance. If the central claim holds, the paper makes a useful contribution by showing that an explicit, staged road-bound goal prior can act as an inductive bias that improves robustness to distribution shift, with only a modest cost on the in-distribution benchmark. The work also has several concrete strengths: the heterogeneous graph with relative edge features is a clean design; the ablation HoliGraph vs. HoliGraph:Goal is an appropriate test of the goal-selection module; the evaluation follows the UniTraj cross-dataset protocol and adds the off-road rate, which is relevant for generalization; and the reported forward-pass time (69 ms) supports real-time applicability. However, the evidence for the generalization claim is currently narrow and lacks statistical support, so the significance is conditional on additional experiments.

major comments (3)
  1. [§IV-C, Table III] The central generalization claim rests on a single transfer direction: training on Argoverse2 and evaluating on NuScenes. The relative-drop improvements over HoliGraph are modest (e.g., b-minFDE6 drops from 10.7% to 8%, minADE6 from 15% to 5.6%), and no reverse transfer (NuScenes to Argoverse2), no held-out-city split within Argoverse2, and no sensitivity analysis for the 2 Hz to 10 Hz linear upscaling described in §IV-A.1 are provided. As written, the reader cannot distinguish a general inductive bias from a property specific to this dataset pair or to the interpolation schema. Please add at least one complementary transfer or held-out evaluation, and report the effect of the upscaling on the NuScenes results.
  2. [Tables I-III] All quantitative results are reported for a single run, with no error bars, confidence intervals, or multiple seeds. Several cross-dataset differences between HoliGraph and HoliGraph:Goal are small in absolute terms (e.g., Table III minFDE6 2.87 m vs. 2.68 m, minMR6 0.41 vs. 0.39), so without variance estimates the claimed improvement could be within run-to-run noise. Please report mean and standard deviation over at least three independent training runs, or provide a statistical significance test for the headline metrics.
  3. [§III-D and §IV-A.3, §IV-C] The near-zero off-road rate (ORR) reported for HoliGraph:Goal is partly mechanical: the goal selection directly chooses lane nodes or point nodes on lanes, and the trajectory is conditioned on a goal point that is regressed from a lane point. Constraining endpoints to the road network by construction will reduce ORR, so comparing ORR with non-goal-based baselines is not an apples-to-apples test of learned generalization. The paper should explicitly acknowledge this and, if ORR is to be used as evidence, report ORR for trajectories before goal refinement or for the goal-regression offset removed, in addition to the current numbers.
minor comments (6)
  1. [Abstract] The sentence beginning 'Latter, is used to classify goals' should read 'The latter is used to classify goals'.
  2. [§IV-A.1] The description of the NuScenes upscaling says 'linear interpolation of 2 consecutive waypoints to fill the resulting gaps of 0.4 s'; it would be clearer to state that this produces 10 Hz sequences and to note whether this matches the UniTraj preprocessing exactly.
  3. [§III-A] In Eq. (1), the time difference is written as $\Delta t_{kl}$ in the equation but as $\Delta t_{mn}$ in the surrounding text; please make the notation consistent.
  4. [§III-D] The term 'accessible lanes' for the (agent-query, decide, lane) edges is not formally defined. Please specify how reachability or accessibility is computed (e.g., by radius, lane connectivity, or drivable-area containment).
  5. [§IV-A.3] The definition of minMRK ('proportion of trajectories with minFDEK > 2 m') is incomplete: it should indicate that the miss is computed per agent and then averaged over agents, and it should clarify the relationship to the NuScenes MissRateTopK_2K definition that uses any-waypoint displacement.
  6. [References] Reference [25] is cited as SceneAttack in §II and §IV-C, but the reference title is 'Scene Transformer: A unified architecture for predicting multiple agent trajectories'; the citation appears to be mismatched with the text.

Circularity Check

0 steps flagged · score 0.0 of 10

Empirical cross-dataset study with a clean ablation baseline; no equation-level or citation-level circularity found.

full rationale

The paper's central claim—that multi-stage goal selection improves cross-dataset generalization—is supported by Table III, which compares HoliGraph:Goal against the authors' own HoliGraph ablation and against Wayformer, Autobot, and MTR numbers taken from UniTraj. No fitted parameter is renamed as a prediction: all quantities in Table III are computed on the held-out NuScenes validation split after training on Argoverse2, so the cross-dataset performance drop is an out-of-sample measurement. The self-citation to the prior HoliGraph architecture [9] is used only as the baseline for the ablation; the goal-selection module is isolated by removing that module, so the comparison is empirical and not a citation-load-bearing argument. The near-zero ORR of HoliGraph:Goal is partly an architectural consequence of binding endpoints to road-network lane points, but this is an inductive-bias and attribution caveat rather than circular equivalence: the model still learns which lane, which point, which offset, and which trajectory to produce, and the ORR is measured, not defined into existence. Potential limitations—the single transfer direction (Av2 to NuScenes), the 2 Hz to 10 Hz linear upscaling of NuScenes, and the absence of reverse transfer—bear on the strength and generality of the generalization claim, not on whether any result reduces to its own inputs. Hence no circularity is found.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The empirical claim rests on the assumption that the relative-drop metric and the 2 Hz to 10 Hz upscaling preserve a meaningful generalization signal, and on several hand-chosen architectural constants (circle numbers, edge radii, loss weights). The model introduces artificial query and point-nrb nodes that are internal components without independent falsifiable evidence.

free parameters (6)
  • point-nrb circle count = 8
    Section III-D: exactly 8 concentric circles are generated around each non-road-bound agent's query node, with radii equal to distance traveled at mean velocity for i seconds. This hand-set number defines the pedestrian goal candidate space and is not tuned by sensitivity analysis.
  • graph edge radii = 125m, 50m, 100m, 150m
    Sections III-B, III-C: lane-to-lane radius 125 m, agent social radius 50 m, lane-to-agent radius 50 m, decoder lane-to-query radius 150 m, and query social radius 100 m. These choices determine which nodes attend to each other and were set without reported tuning.
  • trajectory loss weight = 10
    Section III-F: L = Llane + Lpoint + Lgoal + 10*Ltraj; the factor 10 is hand-set to emphasize trajectory likelihood over goal losses.
  • focal loss alpha and gamma = alpha=0.75, gamma=2
    Section III-F: standard focal loss parameters used as-is; no tuning curve is shown.
  • hidden dimension = 128
    Section IV-A.2: all node and edge features embed to 128 dimensions; this capacity choice is fixed and not ablated.
  • number of graph attention blocks = map-block 1, agent-block 2, query-block 2
    Section III-B and III-C: the map block is executed once, the agent block twice, and two query blocks are stacked; these depths are hand-chosen without ablation.
assumptions (5)
  • domain assumption The HD-map annotations in Argoverse2 and NuScenes can be represented uniformly as lane-nodes and point-nodes with relations successor, predecessor, left-neighbor, right-neighbor, none.
    Section III-A and III-B rely on a homogeneous vectorized map representation; map schema differences between the two datasets (e.g., lane connectivity, point density) could break the staged goal classification.
  • ad hoc to paper Road-bound versus non-road-bound is a sufficient behavioral dichotomy for all traffic participants.
    Section III-D assigns vehicles, trucks, motorcyclists, and cyclists to road-bound and pedestrians to non-road-bound; mixed behaviors (e.g., a cyclist on a sidewalk) are not modeled.
  • domain assumption Linear upsampling of NuScenes from 2 Hz to 10 Hz preserves trajectory shape well enough for metric computation.
    Section IV-A.1 follows UniTraj's interpolation; if interpolation introduces biases, the cross-dataset comparison is affected.
  • domain assumption The relative drop between in-domain and cross-dataset performance is a valid measure of generalization.
    Section IV.C computes 'relative change' as the generalization score; this assumes the two datasets are comparable in difficulty and annotation quality.
  • standard math Transformer-Conv and GNN message passing behave as described in the cited literature.
    Section III-B builds on Masked Label Prediction (Shi et al. 2021); its correctness is assumed from prior publication.
invented entities (2)
  • agent-query-nodes
    purpose: Artificial nodes placed at each agent's last observed position, one per mode, used to pool encoded features and to compute goal scores and trajectory outputs.
    Section III-A defines Q nodes; analogous query mechanisms exist in QC-Net and MTR, but this specific node type is introduced here and has no external falsifiable prediction attached.
  • point-nrb nodes
    purpose: Artificial candidate goal points for non-road-bound agents, arranged on concentric circles with radii based on mean velocity travel distance.
    Section III-D generates these points exclusively inside the model; no outside measurement can confirm or refute their positions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Goal-based Trajectory Prediction for improved Cross-Dataset Generalization." pith.science (2026). https://pith.science/paper/WILBYQ6F

@misc{pith2026250718196,
  author       = {Pith},
  title        = {Pith review of: Goal-based Trajectory Prediction for improved Cross-Dataset Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WILBYQ6F}},
  note         = {Machine review of arXiv:2507.18196}
}
read the original abstract

To achieve full autonomous driving, a good understanding of the surrounding environment is necessary. Especially predicting the future states of other traffic participants imposes a non-trivial challenge. Current SotA-models already show promising results when trained on real datasets (e.g. Argoverse2, NuScenes). Problems arise when these models are deployed to new/unseen areas. Typically, performance drops significantly, indicating that the models lack generalization. In this work, we introduce a new Graph Neural Network (GNN) that utilizes a heterogeneous graph consisting of traffic participants and vectorized road network. Latter, is used to classify goals, i.e. endpoints of the predicted trajectories, in a multi-staged approach, leading to a better generalization to unseen scenarios. We show the effectiveness of the goal selection process via cross-dataset evaluation, i.e. training on Argoverse2 and evaluating on NuScenes.

Figures

Figures reproduced from arXiv: 2507.18196 by the authors.

Figure 1
Figure 1. Schematic illustration of the multi-stage goal selection. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of HoliGraph:Goal. Inputs are embedded in separate embedding modules. The encoder is used to generate [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Architecture of a single Graph Attention layer. [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Qualitative results for cross-evaluation. Models were trained on Argoverse2 [1] and predictions were made an [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages

  1. [1]

    Argoverse 2: Next generation datasets for self-driving perception and forecasting,

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

  2. [2]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 11 621–11 631

  3. [3]

    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 International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 2980–2987

  4. [4]

    Query-centric trajectory prediction,

    Z. Zhou, J. Wang, Y .-H. Li, and Y .-K. Huang, “Query-centric trajectory prediction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 17 863–17 873

  5. [5]

    Unitraj: A unified framework for scalable vehicle trajectory prediction,

    L. Feng, M. Bahari, K. M. B. Amor, É. Zablocki, M. Cord, and A. Alahi, “Unitraj: A unified framework for scalable vehicle trajectory prediction,” in European Conference on Computer Vision . Springer, 2024, pp. 106–123

  6. [6]

    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,” Advances in Neural Information Processing Systems , vol. 35, pp. 6531–6543, 2022

  7. [7]

    HDGT: Heterogeneous Driving Graph Transformer for Multi-Agent Trajectory Prediction via Scene Encoding

    X. Jia, P. Wu, L. Chen, H. Li, Y . Liu, and J. Yan, “Hdgt: Heterogeneous driving graph transformer for multi-agent trajectory prediction via scene encoding,” 2022. [Online]. Available: https://arxiv.org/abs/2205.09753

  8. [8]

    Gorela: Go relative for viewpoint-invariant motion forecasting,

    A. Cui, S. Casas, K. Wong, S. Suo, and R. Urtasun, “Gorela: Go relative for viewpoint-invariant motion forecasting,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 7801–7807

Show all 26 references
  1. [9]

    Holistic graph- based motion prediction,

    D. Grimm, P. Schörner, M. Dreßler, and J.-M. Zöllner, “Holistic graph- based motion prediction,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 2965–2972

  2. [10]

    Covernet: Multimodal behavior prediction using trajectory sets,

    T. Phan-Minh, E. C. Grigore, F. A. Boulton, O. Beijbom, and E. M. Wolff, “Covernet: Multimodal behavior prediction using trajectory sets,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

  3. [11]

    Learning lane graph representations for motion forecasting,

    M. Liang, B. Yang, R. Hu, Y . Chen, R. Liao, S. Feng, and R. Urtasun, “Learning lane graph representations for motion forecasting,” in Computer Vision – ECCV 2020 , A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer International Publishing, 2020, pp. 541–556

  4. [12]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2020

  5. [13]

    Vehicle trajectory prediction works, but not everywhere,

    M. Bahari, S. Saadatnejad, A. Rahimi, M. Shaverdikondori, A. H. Shahidzadeh, S.-M. Moosavi-Dezfooli, and A. Alahi, “Vehicle trajectory prediction works, but not everywhere,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 17 123–17 133

  6. [14]

    Tnt: Target-driven trajectory prediction,

    H. Zhao, J. Gao, T. Lan, C. Sun, B. Sapp, B. Varadarajan, Y . Shen, Y . Shen, Y . Chai, C. Schmid et al. , “Tnt: Target-driven trajectory prediction,” in Conference on Robot Learning . PMLR, 2021, pp. 895–904

  7. [15]

    Gohome: Graph-oriented heatmap output for future motion estimation,

    T. Gilles, S. Sabatini, D. Tsishkou, B. Stanciulescu, and F. Moutarde, “Gohome: Graph-oriented heatmap output for future motion estimation,”

  8. [16]

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

    J. Gu, C. Sun, and H. Zhao, “Densetnt: End-to-end trajectory prediction from dense goal sets,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2021, pp. 15 303– 15 312

  9. [17]

    Masked label prediction: Unified message passing model for semi-supervised classification,

    Y . Shi, Z. Huang, S. Feng, H. Zhong, W. Wang, and Y . Sun, “Masked label prediction: Unified message passing model for semi-supervised classification,” 2021. [Online]. Available: https: //arxiv.org/abs/2009.03509

  10. [18]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garn...

  11. [19]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss for dense object detection,” 2018. [Online]. Available: https://arxiv.org/abs/1708.02002

  12. [20]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101 , 2017

  13. [21]

    Latent variable sequential set transformers for joint multi-agent motion prediction,

    R. Girgis, F. Golemo, F. Codevilla, M. Weiss, J. A. D’Souza, S. E. Kahou, F. Heide, and C. Pal, “Latent variable sequential set transformers for joint multi-agent motion prediction,” 2022. [Online]. Available: https://arxiv.org/abs/2104.00563

  14. [22]

    Multimodal trajectory prediction conditioned on lane-graph traversals,

    N. Deo, E. Wolff, and O. Beijbom, “Multimodal trajectory prediction conditioned on lane-graph traversals,” in Conference on Robot Learning. PMLR, 2022, pp. 203–212

  15. [23]

    Lapred: Lane-aware prediction of multi-modal future trajectories of dynamic agents,

    B. Kim, S. H. Park, S. Lee, E. Khoshimjonov, D. Kum, J. Kim, J. S. Kim, and J. W. Choi, “Lapred: Lane-aware prediction of multi-modal future trajectories of dynamic agents,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 14 636–14 645

  16. [24]

    Thomas: Trajectory heatmap output with learned multi-agent sampling,

    T. Gilles, S. Sabatini, D. Tsishkou, B. Stanciulescu, and F. Moutarde, “Thomas: Trajectory heatmap output with learned multi-agent sampling,” arXiv preprint arXiv:2110.06607 , 2021

  17. [25]

    Scene transformer: A unified architecture for predicting multiple agent trajectories,

    J. Ngiam, B. Caine, V . Vasudevan, Z. Zhang, H.-T. L. Chiang, J. Ling, R. Roelofs, A. Bewley, C. Liu, A. Venugopal, D. Weiss, B. Sapp, Z. Chen, and J. Shlens, “Scene transformer: A unified architecture for predicting multiple agent trajectories,” 2021. [Online]. Available: htt...

  18. [2021]

    Available: https://arxiv.org/abs/2109.01827

    [Online]. Available: https://arxiv.org/abs/2109.01827

Pith tools

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