Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

From Marginal to Joint Predictions: Evaluating Scene-Consistent Trajectory Prediction Approaches for Automated Driving

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 A conditional variational autoencoder formulation beats deterministic decoders on joint trajectory prediction accuracy, though deterministic models still win on interaction robustness.

desk verdict Useful single-backbone comparison of joint prediction approaches, but the central CVAE-vs-deterministic claim is confounded by unequal training epochs and missing error bars. read the letter →

arxiv 2507.05254 v1 pith:OL6B4EOP submitted 2025-07-07 cs.CV cs.AIcs.LGcs.MAcs.RO

Fabian Konstantinidis , Ariel Dallari Guerreiro , Raphael Trumpp , Moritz Sackmann , Ulrich Hofmann , Marco Caccamo , Christoph Stiller This is my paper
classification cs.CVcs.AIcs.LGcs.MAcs.RO
keywords trajectorypredictionjointautomateddrivingconditionalvariationalautoencoderscene-levellossArgoverse2multi-modalitymotionforecasting
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

Predicting every traffic participant jointly, instead of each vehicle independently, is necessary for safe automated driving. This paper takes one strong marginal baseline and adds joint prediction in three ways: recombining marginal modes, training with a scene-level loss (with two decoder upgrades), and reframing the task as a conditional variational autoencoder (CVAE). On the Argoverse 2 test set, every CVAE variant outperforms every deterministic model on the positional metrics minSFDE and minSADE, with the best variant cutting final displacement error by 18.1% and average displacement error by 13.7% relative to marginal recombination. The generative models also produce more plausible and diverse modes, while the deterministic transformer-based model has the lowest collision rate, showing the two families have complementary strengths. The paper's point is that the generative formulation, not just the loss function, is what unlocks accurate scene-consistent forecasts.

What carries the argument

The object that carries the argument is the conditional variational autoencoder built on the same SIMPL backbone. Each actor is assigned a latent token b_i; the posterior network predicts a diagonal Gaussian q(b_i | past, map, ground-truth future), the prior predicts p(b_i | past, map), and a deterministic transformer decoder produces one scene-level trajectory from the sampled latents. Training minimizes L_reg + β·KL(p||q), so the prior learns to generate latent scenes that decode into realistic joint futures. Because the decoder is deterministic, all multimodality is carried by the prior distribution, and the number of predicted modes can be changed by drawing more samples without retraining.

What would settle it

Train the deterministic Joint Loss, Multi-MLP, and Anchor Point Transformer models for 80 epochs with three seeds each on the Argoverse 2 validation set and re-evaluate minSFDE and minSADE; if their values reach or surpass the CVAE numbers, the generative advantage collapses.

Watch

Extended reading notes

Core claim

The central discovery is that how the prediction problem is formulated matters more than merely changing the training objective. Simply replacing the per-agent loss with a scene-level loss degrades accuracy unless the decoder is also made more expressive; the Multi-MLP and Anchor Point Transformer decoders recover and improve performance, and the Anchor Point Transformer attains the lowest actor collision rate (actorCR) of all models. Going further, framing joint prediction as a conditional variational autoencoder, with latent scene variables sampled from a learned prior and fed into a deterministic decoder, yields the best minSFDE and minSADE among all evaluated models, with the Small Beta variant using the prior mean as one mode achieving 2.014 minSFDE and 0.844 minSADE. The generative models show a consistent but slight increase in actorCR, which the authors attribute to sampling from a less constrained prior; the Large Beta variant trades diversity for the lowest actor miss rate among all models.

Load-bearing premise

The comparison assumes that differences in results come from the problem formulation, not from the training budget; however, the generative models were trained for 80 epochs while the deterministic models were trained for 50, and the deterministic models were not rerun for error bars.

Editorial extensions

If this is right

  • A planner supplied with the CVAE Small Beta with mean sampling can expect 18.1% lower final scene displacement error and 13.7% lower average scene displacement error than with marginal recombination on Argoverse 2.
  • Replacing the per-agent loss with a scene-level loss alone hurts accuracy; the gains only appear when the decoder is restructured, and the Anchor Point Transformer version yields the lowest actor collision rate.
  • Because multimodality in the CVAE comes from sampling the latent prior, the same trained model can produce any number of modes at inference time, trading latency against coverage without retraining.
  • The generative and deterministic families have complementary strengths: generative models win on positional accuracy and mode plausibility, while deterministic models show more robust interaction quality as measured by actorCR.
  • Using the prior mean as one mode gives the best overall balance, suggesting a practical way to get a deterministic best guess out of a generative model.

Reading between the lines

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

  • If the extra 30 training epochs given only to the CVAE models explain part of the gap, then retraining the deterministic decoders for 80 epochs with repeated seeds would narrow or close the 18.1% difference; the paper does not run this control.
  • The consistently higher actorCR of generative models suggests a practical safeguard: sample many CVAE scenes and filter them with a deterministic interaction scorer before passing them to a planner, a hybrid the paper does not evaluate.
  • The authors note that the CVAE latents are actor-specific but scene-informed; a shared latent scene variable, which they list as future work, would be a natural test of whether actor-specific latents are what preserve scene consistency.
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 a systematic comparison of approaches to joint trajectory prediction for automated driving, all built on the SIMPL marginal baseline. The evaluated strategies are (1) recombination of marginal modes, (2) scene-level loss training with three decoder variants (Joint Loss, Multi-MLP, Anchor Point Transformer), and (3) a conditional variational autoencoder (CVAE) formulation with two beta settings and two sampling variants. Using the Argoverse 2 Motion Forecasting test set, the authors report minSFDE, minSADE, actorMR, actorCR, and inference times, and they conclude that CVAE-based models achieve superior positional accuracy while deterministic models exhibit lower actorCR. A qualitative multi-modality analysis of one selected scene and a project webpage supplement the quantitative evaluation.

Significance. If the conclusions hold, the paper provides a practically useful controlled comparison of joint prediction formulations, which is genuinely needed because prior work conflates architecture, training, and problem formulation. The paper's strength is that all methods share a common SIMPL backbone, making the comparison more controlled than typical leaderboard evaluations. The CVAE models were evaluated with three submissions, allowing error bars on the headline numbers. However, the central quantitative claim is weakened by a training-budget confound, and the interaction metrics are not formally defined, so the current evidence does not fully support the stated conclusions.

major comments (3)
  1. [Section V, Training Details and Table I] The headline claim that 'All CVAE-based models outperform all deterministic models in terms of the positional errors minSFDE and minSADE' is not supported as an attribution to the generative formulation because the CVAE models were trained for 80 epochs while deterministic models were trained for 50. The text states that CVAE training was extended by 30 epochs 'to ensure fair comparison,' but no equivalent budget was given to deterministic models, and no repeated-seed statistics are reported for the deterministic rows. The minSFDE gap between the best CVAE variant (2.014) and the best deterministic model (2.197) could therefore reflect additional optimization rather than the generative formulation. Please re-run deterministic models for 80 epochs (or to their own convergence criterion), run all models with multiple seeds, and report means and variances for every row in Table I, or explicitly restrict the conclusion to the specific training budgets used.
  2. [Section V, Evaluation Metrics] The metrics actorCR and actorMR are mentioned but never formally defined. The text says only that they are 'auxiliary metrics evaluating the intra-mode consistency, such as Actor Collision Rate and Actor Miss Rate,' which is insufficient for a paper whose secondary claim is that deterministic models 'show lower actorCR, indicating more robust interaction quality.' Please provide precise definitions: what constitutes a collision, how the miss rate threshold is set, whether these metrics are computed per predicted mode or per scene, and how they aggregate over the K modes.
  3. [Section V-B, Multi-modality] The qualitative multi-modality analysis is based on a single selected scene (Figure 3). The claims that deterministic models 'often produce implausible modes' while the CVAE model 'produces more plausible and genuinely multi-modal predictions' are anecdotal when supported by only one example. Please either provide quantitative diversity/plausibility metrics over the full test set or analyze multiple representative scenes, and state how the scene was selected.
minor comments (5)
  1. [Section IV-B, Eq. (5)] For the CVAE models, K=1 in the regression loss, so the min over k is vacuous; please clarify that the scene-level WTA loss reduces to a single-mode regression loss in the generative case.
  2. [Table I and Section IV-C] The footnote indicating that the prior mean is used for one mode is helpful, but please clarify how this affects the interpretation of minSFDE and minSADE: are all six modes still counted, and is the mean-conditioned mode always included in the min computation?
  3. [Section V-C, Table II] The linear regression coefficients γ_a and γ_l are reported without confidence intervals or significance levels; please add these or state that the values are only descriptive.
  4. [Throughout] There are minor typographical issues, including 'Kullback Leiber' (should be 'Kullback-Leibler'), inconsistent spacing in the abstract ('CV AE' vs 'CVAE'), and the use of 'CV AE' and 'CVAE' interchangeably; please unify terminology.
  5. [Figure 3] The figure would benefit from labels indicating which vehicle is the 'purple' and which is the 'blue' vehicle referenced in the text, and from consistent color rendering for readers with color vision deficiencies.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: the paper is an empirical benchmark comparison; the only self-citation (SceneMotion) is architectural inspiration, not a load-bearing premise.

full rationale

This paper does not derive its prediction results from its own definitions. The central claim that CVAE-based models outperform deterministic models on minSFDE and minSADE is an empirical result from Table I, obtained by training each model on Argoverse 2 and evaluating on the competition test set. No equation in the paper defines the reported metric in terms of the model family, and no fitted parameter is later renamed as a prediction. The marginal-recombination confidence score in Eq. (4) and the scene-level loss in Eq. (5) are alternative formulations under comparison, not assumed conclusions. The only self-citation to prior work by the same authors is SceneMotion [9], used as a design reference for the anchor-point transformer and as one example of scene-level-loss training; the paper's argument does not reduce to that citation, and no uniqueness claim is imported. The methodological concern about unequal training budgets—CVAE models received 80 epochs while deterministic models received 50, justified by the statement 'Since the two CV AE-based models had not yet converged after 50 epochs, as indicated by their validation loss, we extended their training by an additional 30 epochs to ensure fair comparison'—is a potential confound in the Table I comparison, but it is a fairness and correctness issue, not circularity. The training budget is not defined in terms of the test metrics, and the comparison is not tautological. Accordingly, no circular step meeting the required evidentiary standard is present.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The central claim is an empirical comparison, so the ledger captures the main hyperparameters and modeling assumptions on which the conclusions rest. No new physical entities are introduced; the latent scene variables are standard CVAE machinery, not independently evidenced entities.

free parameters (6)
  • beta (KL weight) = 0.05 (small) and 0.5 (large)
    Controls the trade-off between reconstruction and KL divergence in the CVAE loss (Eq. 6); chosen manually, directly affects diversity and accuracy.
  • latent scene dimension DB = 32
    Selected by ablation on a 20% data subset (evaluated against 64); affects the quality of sampled latent scenes.
  • Anchor Point Transformer layers = 2
    Selected by ablation on a 20% subset (evaluated against 1 and 3); impacts inter-mode coordination and inference time.
  • training epochs = 50 deterministic, 80 CVAE
    CVAE models were given 30 additional epochs because they had not converged, creating a confounding factor in the comparison.
  • number of SFT layers = 4 total per model (CVAE: 2 in prior, 2 in decoder)
    Chosen to keep inference-time consistency across models; a design choice affecting representational capacity.
  • hidden dimension D = 128
    Common hidden size for all models, set without ablation.
assumptions (7)
  • domain assumption Smooth L1 regression loss and cross-entropy confidence loss are appropriate for trajectory prediction.
    Used in Eq. (5); standard practice in the field, not independently justified in this paper.
  • domain assumption Gaussian posterior and prior with diagonal covariance adequately capture the latent scene distribution.
    Follows ref. [11]; a modeling choice that could limit multimodal diversity.
  • domain assumption Argoverse 2 ground-truth trajectories and competition metrics are a valid proxy for real-world prediction quality.
    The evaluation is entirely based on this dataset; performance on it is treated as the ground truth for comparing methods.
  • ad hoc to paper The instance-centric SIMPL backbone transfers to joint prediction without architectural changes beyond the decoder.
    The paper builds all joint models on SIMPL encoders; if the backbone is a bottleneck for joint reasoning, the comparison is limited.
  • domain assumption Winner-Takes-All loss causes implausible modes that are not penalized, motivating the generative formulation.
    Stated in Sections IV-C and V-B; a hypothesis supported by one qualitative example.
  • domain assumption Recombination via the product of marginal confidence scores identifies good joint scenes.
    Eq. (4), borrowed from MTR++/M2I; assumes confidence scores are comparable across agents and modes.
  • standard math The ELBO objective (Eq. 6) with beta weighting is a valid training objective for scene prediction.
    Standard variational inference machinery, used in ref. [11].

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Marginal to Joint Predictions: Evaluating Scene-Consistent Trajectory Prediction Approaches for Automated Driving." pith.science (2026). https://pith.science/paper/OL6B4EOP

@misc{pith2026250705254,
  author       = {Pith},
  title        = {Pith review of: From Marginal to Joint Predictions: Evaluating Scene-Consistent Trajectory Prediction Approaches for Automated Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OL6B4EOP}},
  note         = {Machine review of arXiv:2507.05254}
}
read the original abstract

Accurate motion prediction of surrounding traffic participants is crucial for the safe and efficient operation of automated vehicles in dynamic environments. Marginal prediction models commonly forecast each agent's future trajectories independently, often leading to sub-optimal planning decisions for an automated vehicle. In contrast, joint prediction models explicitly account for the interactions between agents, yielding socially and physically consistent predictions on a scene level. However, existing approaches differ not only in their problem formulation but also in the model architectures and implementation details used, making it difficult to compare them. In this work, we systematically investigate different approaches to joint motion prediction, including post-processing of the marginal predictions, explicitly training the model for joint predictions, and framing the problem as a generative task. We evaluate each approach in terms of prediction accuracy, multi-modality, and inference efficiency, offering a comprehensive analysis of the strengths and limitations of each approach. Several prediction examples are available at https://frommarginaltojointpred.github.io/.

Figures

Figures reproduced from arXiv: 2507.05254 by the authors.

Figure 1
Figure 1. Comparison of marginal predictions and joint predictions for [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed CVAE architecture, composed of the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Scene-level multi-modal predictions from each model, along with their corresponding SFDE and SADE scores. Past and predicted future trajectories [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Long-term Traffic Scene Prediction via Polynomial Representations in Autonomous Driving

    cs.AI 2026-08 conditional novelty 6.0 of 10

    Polynomial representations of trajectories and maps yield competitive prediction accuracy while substantially improving cross-dataset generalization and computational efficiency in autonomous driving.

Reference graph

Works this paper leans on

32 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Marginal Recombination (Section IV-A): The marginal baseline model evaluated jointly by recombining pre- dicted marginal modes

    work page

  2. [2]

    Joint Loss (Section IV-B): The baseline model trained directly with a scene-level loss to encourage joint consistency

    work page

  3. [3]

    Multi-MLP (Section IV-B): An extension of the baseline with a Multi-MLP decoder, using separate heads for each mode, trained with the scene-level loss

    work page

  4. [4]

    Anchor Point Transformer (Section IV-B): An extension of the baseline with a Transformer-based decoder to improve inter-mode accuracy, also trained with the scene-level loss

    work page

  5. [5]

    CVAE (Section IV-C): The baseline model adapted to the CV AE framework, enabling sampling-based generation of predictions. A. Recombination of Marginal Modes Our first approach for extending the baseline model to joint predictions does not involve any retraining. Instead, following the method proposed in [5], [6], we simply recombine the marginal predicti...

    work page

  6. [6]

    Simpl: A simple and efficient multi-agent motion prediction baseline for autonomous driving,

    L. Zhang, P. Li, S. Liu, and S. Shen, “Simpl: A simple and efficient multi-agent motion prediction baseline for autonomous driving,” IEEE Robotics and Automation Letters , 2024

    2024

  7. [7]

    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 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2020, pp. 11 522–11 530

    work page 2020

  8. [8]

    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 , 2021, pp. 15 303–15 312

    2021

Show all 32 references
  1. [9]

    The integration of prediction and planning in deep learning automated driving systems: A review,

    S. Hagedorn, M. Hallgarten, M. Stoll, and A. P. Condurache, “The integration of prediction and planning in deep learning automated driving systems: A review,” IEEE Transactions on Intelligent V ehicles , 2024

    2024

  2. [10]

    Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying,

    S. Shi, L. Jiang, D. Dai, and B. Schiele, “Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  3. [11]

    M2i: From factored marginal trajectory prediction to interactive prediction,

    Q. Sun, X. Huang, J. Gu, B. C. Williams, and H. Zhao, “M2i: From factored marginal trajectory prediction to interactive prediction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 6543–6552

    2022

  4. [12]

    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

    2021 arXiv

  5. [13]

    Scene transformer: A unified architecture for predicting future trajectories of multiple agents,

    J. Ngiam, V . Vasudevan, B. Caine, Z. Zhang, H.-T. L. Chiang, J. Ling, R. Roelofs, A. Bewley, C. Liu, A. Venugopal, D. J. Weiss, B. Sapp, Z. Chen, and J. Shlens, “Scene transformer: A unified architecture for predicting future trajectories of multiple agents,” in International...

    2022

  6. [14]

    Scenemotion: From agent-centric embeddings to scene-wide forecasts,

    R. Wagner, ¨Omer Sahin Tas, M. Steiner, F. Konstantinidis, H. K ¨onigshof, M. Klemp, C. Fernandez, and C. Stiller, “Scenemotion: From agent-centric embeddings to scene-wide forecasts,” 2024. [Online]. Available: https://arxiv.org/abs/2408.01537

    2024 arXiv

  7. [15]

    MotionLM: Multi-Agent Motion Forecasting as Language Modeling ,

    A. Seff, B. Cera, D. Chen, M. Ng, A. Zhou, N. Nayakanti, K. S. Refaat, R. Al-Rfou, and B. Sapp, “ MotionLM: Multi-Agent Motion Forecasting as Language Modeling ,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV) . Los Alamitos, CA, USA: IEEE Computer Society...

    2023

  8. [16]

    Implicit latent variable model for scene-consistent motion forecasting,

    S. Casas, C. Gulino, S. Suo, K. Luo, R. Liao, and R. Urtasun, “Implicit latent variable model for scene-consistent motion forecasting,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIII 16 . Springer, 2020, pp. 624–641

    2020

  9. [17]

    Lookout: Diverse multi-future prediction and planning for self-driving,

    A. Cui, S. Casas, A. Sadat, R. Liao, and R. Urtasun, “Lookout: Diverse multi-future prediction and planning for self-driving,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, Oct. 2021, p. 16087–16096. [Online]. Available: http://dx.doi.org/10.1109/I...

    2021

  10. [18]

    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, D. Ramanan, P. Carr, and J. Hays, “Argoverse 2: Next generation datasets for self-driving perception and forecasting,” in Proceedings of the Neural Information Proce...

    2021

  11. [19]

    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) , 2023, pp. 2980–2987

    2023

  12. [20]

    Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding,

    Z. Zhang, A. Liniger, C. Sakaridis, F. Yu, and L. Van Gool, “Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding,” in Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  13. [21]

    Trajeglish: Traffic modeling as next-token prediction,

    J. Philion, X. B. Peng, and S. Fidler, “Trajeglish: Traffic modeling as next-token prediction,” in The Twelfth International Conference on Learning Representations , 2024. [Online]. Available: https://openreview.net/forum?id=Z59Rb5bPPP

    2024

  14. [22]

    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 (CVPR) , 2023

    2023

  15. [23]

    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 Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023

    2023

  16. [24]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 2039–2049

    2024

  17. [25]

    Multimodal motion prediction with stacked transformers,

    Y . Liu, J. Zhang, L. Fang, Q. Jiang, and B. Zhou, “Multimodal motion prediction with stacked transformers,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2021, pp. 7573– 7582

    2021

  18. [26]

    San: Scene anchor networks for joint action-space prediction,

    F. Janjoˇs, M. Dolgov, M. Kuri ´c, Y . Shen, and J. M. Z ¨ollner, “San: Scene anchor networks for joint action-space prediction,” in 2022 IEEE Intelligent V ehicles Symposium (IV) . IEEE, 2022, pp. 1751–1756

    2022

  19. [27]

    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 ECCV, 2020

    2020

  20. [28]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 652–660

    2017

  21. [29]

    Fast r-cnn,

    R. Girshick, “Fast r-cnn,” in 2015 IEEE International Conference on Computer Vision (ICCV) , 2015, pp. 1440–1448

    2015

  22. [30]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

    2017

  23. [31]

    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 International Conference on Robotics ...

    2022

  24. [32]

    Hierarchical latent structure for multi-modal vehicle trajectory forecasting,

    D. Choi and K. Min, “Hierarchical latent structure for multi-modal vehicle trajectory forecasting,” in European conference on computer vision. Springer, 2022, pp. 129–145

    2022

Pith tools

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