Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

TopoDiffuser: A Diffusion-Based Multimodal Trajectory Prediction Model with Topometric Maps

T0 review · 3 major / 4 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 TopoDiffuser claims that conditioning a diffusion model on a topometric map—a binary OpenStreetMap route corridor—produces future trajectories that are both more accurate and more road-compliant than prior methods, without explicit hard…

desk verdict The headline accuracy gains are almost certainly an artifact of conditioning on the future path: the OSM route includes 15 future keyframes and the road-segmentation target is the recorded trajectory, so the comparison against baselines is invalid. read the letter →

arxiv 2508.00303 v1 pith:CMPCUNG7 submitted 2025-08-01 cs.RO

Zehui Xu , Junhui Wang , Yongliang Shi , Chao Gao , Guyue Zhou This is my paper
classification cs.RO
keywords trajectorypredictiondiffusionmodelstopometricmapsmultimodalfusionbird's-eye-viewautonomousdrivingKITTIbenchmarkroadcompliance
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

TopoDiffuser claims that feeding a topometric map—here a binary OpenStreetMap route corridor—into a conditional diffusion model makes predicted future trajectories both more accurate and more road-compliant, without needing explicit hard constraints. On the KITTI raw dataset it reports final displacement errors of 0.56 m, 0.31 m, and 0.46 m on sequences 08, 09, and 10, beating the next-best baseline by 33–44% in FDE on 08 and 09. The model fuses LiDAR, ego-history, and the map route into one bird's-eye-view tensor, learns a road segmentation mask as conditioning, and samples multiple denoised trajectories at inference. If the result holds, it would mean diffusion-based prediction can combine multimodal diversity with geometric feasibility using only coarse, off-the-shelf map data.

What carries the argument

The central object is a conditional denoising diffusion model whose condition is a compact vector produced by a multimodal conditioning encoder. The encoder turns $\text{Concat}(I_{\text{lidar}}, I_{\text{traj}}, I_{\text{map}})$—three LiDAR BEV channels plus binary history and OSM-route masks—into a feature map, predicts a road segmentation mask as an auxiliary task, and reshapes the mask features into the conditioning vector $c$. A lightweight U-Net $g_\varphi(\tau_t, t, c)$ predicts the noise at each of ten denoising steps, and the auxiliary road segmentation is trained against a mask made by rasterizing the recorded driving trajectory. This design is what lets the map influence generation softly rather than through explicit constraints.

What would settle it

Re-run the KITTI experiments with the OSM route clipped to history-only or to a destination fixed before the future is known, and replace the recorded-trajectory road mask with an independently sourced road map; if the FDE and minADE gains over MTP disappear, the reported improvements come from future-route leakage rather than learned road geometry.

Watch

Extended reading notes

Core claim

The paper's central claim is that structural cues from topometric maps can be embedded directly into the reverse diffusion process, so that generated trajectories naturally stay on the road while still capturing multiple plausible futures. Concretely, TopoDiffuser concatenates a LiDAR BEV encoding, a rasterized history mask, and an OSM-derived route mask into a five-channel input, processes it with a CNN that also predicts a drivable-area segmentation, and conditions a small denoising U-Net on the resulting vector. Trained with a noise-prediction MSE loss plus a road-segmentation cross-entropy term, and sampling five trajectories at inference, it reports state-of-the-art numbers on KITTI sequences 08, 09, and 10, including a Hausdorff distance of 1.33 m on 08, which the authors take as evidence of strong geometric consistency.

Load-bearing premise

The route input to the model includes 15 keyframes into the future, and the road mask used for supervision is the rasterized recorded driving trajectory; if those count as handing the model the future path, the evaluation is not a fair test of forecasting ability.

Editorial extensions

If this is right

  • If the central claim holds, diffusion-based predictors can achieve road compliance without hard constraints, simply by conditioning on coarse route masks.
  • The reported 33–44% FDE improvements on KITTI-08/09 imply that topometric conditioning is a cheap and effective substitute for heavier HD-map or lane-graph encoders.
  • Because more denoising steps (up to about 20) and more samples (up to about 8) monotonically improve metrics before saturating, deployment can trade compute for accuracy in safety-critical settings.
  • The ablations imply each modality contributes: the map alone improves minADE and Hausdorff distance, while adding history further reduces Hausdorff distance, so the fusion scheme is what carries the gain.

Reading between the lines

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

  • The 15 future keyframes in the OSM route mean part of the reported gain may be route information leakage; a version that uses only past-available map data would clarify how much genuine forecasting skill remains.
  • Because the road-segmentation ground truth is the recorded trajectory itself, the auxiliary head could be learning to copy the driven path; testing with independent road labels would separate map understanding from trajectory memorization.
  • The same conditioning scheme might transfer to predicting other agents or to map sources other than OSM, but the paper does not demonstrate that, so it remains an open extension.
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 / 4 minor

Summary. TopoDiffuser proposes a conditional diffusion model for multimodal ego-trajectory prediction on KITTI. The input is a BEV tensor formed by concatenating LiDAR features, rasterized historical ego positions, and an OSM-derived binary route mask; a CNN conditioning encoder also predicts an auxiliary road segmentation mask during training, and a lightweight U-Net denoises future trajectories conditioned on the resulting context vector. The paper reports state-of-the-art results on KITTI sequences 08, 09, and 10 (e.g., FDE of 0.56 m on KITTI-08, 0.31 m on KITTI-09), together with ablations on input modalities, denoising steps, and sample count. The code is publicly released.

Significance. If the empirical claims were valid, the paper would offer a useful application of diffusion models to trajectory prediction with map guidance, and the open-source release would aid reproducibility. However, the central comparison is not a valid test of the proposed method: the OSM route conditioning input explicitly contains the future driving path (15 future keyframes, Section IV.B), and the auxiliary road-segmentation supervision is the rasterized recorded driving trajectory (Section III.E). The reported improvements over baselines that receive no such future-route channel are therefore not interpretable as evidence for topometric-map-guided prediction. The paper does provide a clearly stated problem formulation, a reproducible training objective, and several ablations, but the load-bearing evaluation protocol invalidates the headline claim. Consequently, the contribution is not established in its current form.

major comments (3)
  1. [IV.B and Eq. (2)] The OSM route is defined as "centered at the current ego position and covering both the past and future driving paths," with 5 keyframes into the past and 15 into the future, sampled every 2 meters. This is not a navigational route prior; it is a sampled trace of the ground-truth future trajectory. Because this mask is concatenated into the input tensor in Eq. (2) and processed by the conditioning encoder, the model observes a corridor from which the future endpoint and general shape are largely recoverable. None of the baselines (CoverNet, MTP, TP) receive a comparable future-route channel, so the improvements in Table I (e.g., 33-44% FDE reduction over MTP) are attributable to information leakage rather than to topometric map guidance. This is a fundamental flaw in the experimental design.
  2. [III.E and Eq. (4)] The auxiliary road-segmentation head is supervised with a ground-truth mask constructed by rasterizing the recorded driving trajectory. Although the head is discarded at inference, the shared CNN backbone is jointly trained with this future-informed signal (L_total = L_diffusion + lambda_road * L_road). This means the encoder learns features that encode the future path during training, further compounding the leakage identified above and preventing any clean interpretation of the road-compliance or accuracy gains.
  3. [Table I and Section IV.D] All quantitative results are reported as single point estimates with no error bars, no repeated runs, and no variance information. Given that the method samples 5 trajectories with independent Gaussian noise, performance is stochastic; without multiple seeds or confidence intervals, the reported improvements (many of which are at the centimeter level) cannot be distinguished from sampling noise. This is especially problematic because the main comparison is already confounded by the future-route input.
minor comments (4)
  1. [Eq. (6)] The FDE formula is missing the explicit norm: it should be written as ||s_hat_{T-1} - s*_{T-1}||_2 rather than the implied double bars.
  2. [IV.D.3 and Table II] The ablation study reports only KITTI-10 and only one configuration per row; it would be helpful to report variances and to state clearly whether the "L." model still receives the OSM route mask, since the route channel is the primary source of suspected leakage.
  3. [II.C and IV.D.1] The baseline implementations are not fully specified; for example, CoverNet requires a predefined trajectory set whose size and sampling strategy are not given for this KITTI setup, making the comparison hard to reproduce.
  4. [Fig. 1] Figure 1 is referenced in the method section but not described in enough detail; in particular, the diagram omits the auxiliary segmentation head and the training-time losses, which would clarify the architecture.

Circularity Check

2 steps flagged · score 8.0 of 10

Central benchmark gains are forced by feeding the future trajectory into the conditioning input: the OSM route covers 15 future keyframes and the road-segmentation target is the rasterized recorded driving path.

  1. self definitional [Section IV.B (Implementation Details) with Eq. (2) in Section III.B]
    "For topological guidance, we extract a route from OSM, centered at the current ego position and covering both the past and future driving paths. Specifically, the OSM route includes 5 keyframes into the past and 15 into the future, sampled every 2 meters."

    The input tensor defined in Eq. (2) is Iinput = Concat(Ilidar, Itraj, Imap), where Imap is the binary mask of this route. The route is not an independent map prior; it is sampled from the future driving path (15 keyframes ahead), i.e., from the very trajectory tau that the model is supposed to predict. Conditioning the reverse diffusion process on c derived from this mask hands the model the corridor containing the ground-truth future path, so the reported FDE/minADE improvements over baselines that receive no such channel are largely a property of the input, not of learned prediction. The 'prediction' reduces, by construction, to following the provided future corridor.

  2. other [Section III.E (Training and Inference), Eq. (4), with Section III.C]
    "The ground-truth segmentation mask is constructed by rasterizing the recorded driving trajectory into a binary image y in {0,1}^{H'xW'}, where each pixel indicates whether it belongs to the traversed road region."

    This ground-truth mask y is the future trajectory itself. The auxiliary head is trained with Lroad to predict y, and Section III.C states that the same CNN backbone 'simultaneously predict[s] a road segmentation mask' before producing the conditioning vector c. The backbone is therefore explicitly optimized to recognize the future driving path from the input tensor. Although the head is discarded at inference, the shared encoder has been trained to encode the future route into c, and c conditions the denoising network g_phi at every step. This creates a second, training-time channel through which the target future trajectory is embedded in the conditioning signal, so the road-compliance and accuracy claims are not independent of the target.

full rationale

The paper's central empirical claim—that TopoDiffuser outperforms state-of-the-art methods in FDE, minADE, HitRate, and HD on KITTI—is undermined by two coupled construction choices. First, the topometric 'route' input is a sampled version of the future driving path (5 keyframes past, 15 future, every 2 m) and is concatenated into the conditioning tensor (Eq. 2). This is not a future-free navigational prior; it is the ground-truth trajectory expressed as a drivable corridor. Second, the auxiliary road-segmentation supervision uses the rasterized recorded driving trajectory as its ground truth (Eq. 4), and the conditioning encoder shares the backbone with that segmentation task, so the model is trained to encode the future path into the context vector c. Both channels are absent in the CoverNet, MTP, and TP baselines, making the Table I comparison informationally unequal. The magnitude of the reported advantage (e.g., FDE 0.56 m on KITTI-08) and the ablation improvements attributed to map guidance therefore reduce, by construction, to the model's access to the future trajectory rather than to a learned topometric prior. This is not a case of a benign self-citation; it is a self-definitional leakage of the prediction target into the conditioning inputs. Score 8 rather than 10 because the diffusion decoder still performs some nontrivial regression, but the core accuracy comparison is structurally forced.

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

No new physical entities are introduced, but two design choices create leakage: the future route appears in the map input, and the road mask is derived from the driven trajectory.

free parameters (4)
  • lambda_road = not specified
    Loss weighting for road segmentation, balances L_diffusion and L_road.
  • noise schedule gamma_t = not specified
    Monotonically decreasing schedule used in forward diffusion, not fully quantified.
  • denoising steps = 10
    Number of reverse diffusion steps; ablation shows performance varies with this.
  • number of samples = 5
    Inference-time trajectory samples; ablation shows metrics improve with more samples.
assumptions (3)
  • domain assumption OSM route covering both past and future driving paths is a valid topometric map prior.
    Section III.B explicitly includes future keyframes, conflating map prior with ground-truth future path.
  • domain assumption The recorded driving trajectory rasterized into a binary mask is a valid ground-truth road segmentation.
    Section III.E uses the traversed road region as supervision for the segmentation head, leaking future trajectory.
  • standard math Standard diffusion and CNN backbones operate correctly.
    The conditional diffusion formulation follows standard practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TopoDiffuser: A Diffusion-Based Multimodal Trajectory Prediction Model with Topometric Maps." pith.science (2026). https://pith.science/paper/CMPCUNG7

@misc{pith2026250800303,
  author       = {Pith},
  title        = {Pith review of: TopoDiffuser: A Diffusion-Based Multimodal Trajectory Prediction Model with Topometric Maps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMPCUNG7}},
  note         = {Machine review of arXiv:2508.00303}
}
read the original abstract

This paper introduces TopoDiffuser, a diffusion-based framework for multimodal trajectory prediction that incorporates topometric maps to generate accurate, diverse, and road-compliant future motion forecasts. By embedding structural cues from topometric maps into the denoising process of a conditional diffusion model, the proposed approach enables trajectory generation that naturally adheres to road geometry without relying on explicit constraints. A multimodal conditioning encoder fuses LiDAR observations, historical motion, and route information into a unified bird's-eye-view (BEV) representation. Extensive experiments on the KITTI benchmark demonstrate that TopoDiffuser outperforms state-of-the-art methods, while maintaining strong geometric consistency. Ablation studies further validate the contribution of each input modality, as well as the impact of denoising steps and the number of trajectory samples. To support future research, we publicly release our code at https://github.com/EI-Nav/TopoDiffuser.

Figures

Figures reproduced from arXiv: 2508.00303 by the authors.

Figure 1
Figure 1. Overview of the proposed diffusion-based trajectory prediction [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Predicted trajectories on representative scenes. Blue for OSM route, red for ground truth, green for predictions. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Effect of number of sampled trajectories on prediction accuracy. [PITH_FULL_IMAGE:figures/full_fig_p005_4.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. SARe: Structure-Aware Generative 3D Fragment Reassembly

    cs.CV 2026-03 conditional novelty 6.0 of 10

    SARe improves many-fragment 3D reassembly by jointly predicting fracture-surface labels and a contact graph during flow-based pose generation, plus inference-time resampling of uncertain regions.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    Argoverse: 3d tracking and forecasting with rich maps,

    M.-F. Chang, J. Lambert, P. Sangkloy, J. Singh, S. Bak, A. Hartnett, D. Wang, P. Carr, S. Lucey, D. Ramanan, and J. Hays, “Argoverse: 3d tracking and forecasting with rich maps,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , June 2019

    work page 2019

  2. [2]

    Multi-modal trajectory prediction of surrounding vehicles with maneuver-based lstms,

    N. Deo and M. M. Trivedi, “Multi-modal trajectory prediction of surrounding vehicles with maneuver-based lstms,” inProc. IEEE Intell. V ehicles Symp. (IV), 2018, pp. 1179–1184

    work page 2018

  3. [3]

    Predicting vehicle behaviors over an extended horizon using behavior interaction network,

    W. Ding, J. Chen, and S. Shen, “Predicting vehicle behaviors over an extended horizon using behavior interaction network,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA) , 2019, pp. 8634–8640

    work page 2019

  4. [4]

    Deeptrack: Lightweight deep learning for vehicle trajectory prediction in highways,

    V . Katariya, M. Baharani, N. Morris, O. Shoghli, and H. Tabkhi, “Deeptrack: Lightweight deep learning for vehicle trajectory prediction in highways,” IEEE Trans. Intell. Transp. Syst. , vol. 23, no. 10, pp. 18 927–18 936, 2022

    work page 2022

  5. [5]

    Vehicle trajectory prediction for automated driving based on temporal convolution networks,

    D. Li, H. Li, Y . Xiao, B. Li, and B. Tang, “Vehicle trajectory prediction for automated driving based on temporal convolution networks,” in Proc. WRC Symp. Adv. Robot. Autom. (WRC SARA) , 2022, pp. 257– 262

    work page 2022

  6. [6]

    Long-term prediction for high-resolution lane-changing data using temporal convolution network,

    Y . Zhang, Y . Zou, J. Tang, and J. Liang, “Long-term prediction for high-resolution lane-changing data using temporal convolution network,” Transportmetrica B: Transport Dyn. , vol. 10, no. 1, pp. 849–863, Jul. 2021

    work page 2021

  7. [7]

    S2tnet: Spatio-temporal transformer networks for trajectory prediction in autonomous driving,

    W. Chen, F. Wang, and H. Sun, “S2tnet: Spatio-temporal transformer networks for trajectory prediction in autonomous driving,” in Proc. 13th Asian Conf. Mach. Learn. (ACML) , vol. 157. PMLR, Nov. 17–19 2021, pp. 454–469

    work page 2021

  8. [8]

    Multimodal motion prediction with stacked transformers,

    Y . Liu, J. Zhang, L. Fang, Q. Jiang, and B. Zhou, “Multimodal motion prediction with stacked transformers,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , June 2021, pp. 7577– 7586

    work page 2021

Show all 30 references
  1. [9]

    Predicting vehicles trajectories in urban scenarios with trans- former networks and augmented information,

    A. Quintanar, D. Fern ´andez-Llorca, I. Parra, R. Izquierdo, and M. A. Sotelo, “Predicting vehicles trajectories in urban scenarios with trans- former networks and augmented information,” in Proc. IEEE Intell. V ehicles Symp. (IV), 2021, pp. 1051–1056

    2021

  2. [10]

    Forecasting trajectory and behavior of road-agents using spectral clustering in graph-lstms,

    R. Chandra, T. Guan, S. Panuganti, T. Mittal, U. Bhattacharya, A. Bera, and D. Manocha, “Forecasting trajectory and behavior of road-agents using spectral clustering in graph-lstms,” IEEE Robot. Autom. Lett. , vol. 5, no. 3, pp. 4882–4890, 2020

    2020

  3. [11]

    Grip: Graph-based interaction- aware trajectory prediction,

    X. Li, X. Ying, and M. C. Chuah, “Grip: Graph-based interaction- aware trajectory prediction,” in Proc. IEEE Intell. Transp. Syst. Conf. (ITSC), 2019, pp. 3960–3966

    2019

  4. [12]

    Trajectory prediction-based local spatio-temporal navigation map for autonomous driving in dynamic highway environments,

    M. Fu, T. Zhang, W. Song, Y . Yang, and M. Wang, “Trajectory prediction-based local spatio-temporal navigation map for autonomous driving in dynamic highway environments,” IEEE Trans. Intell. Transp. Syst., vol. 23, no. 7, pp. 6418–6429, 2022

    2022

  5. [13]

    Intelligent vehicles trajectory prediction with spatial and temporal attention mechanism,

    Q. Meng, B. Shang, Y . Liu, H. Guo, and X. Zhao, “Intelligent vehicles trajectory prediction with spatial and temporal attention mechanism,” IF AC-PapersOnLine, vol. 54, no. 10, pp. 454–459, 2021

    2021

  6. [14]

    Map- enhanced generative adversarial trajectory prediction method for au- tomated vehicles,

    H. Guo, Q. Meng, X. Zhao, J. Liu, D. Cao, and H. Chen, “Map- enhanced generative adversarial trajectory prediction method for au- tomated vehicles,” Inf. Sci. , vol. 622, pp. 1033–1049, 2023

    2023

  7. [15]

    Vehicle trajectory prediction using gan,

    C. Hegde, S. Dash, and P. Agarwal, “Vehicle trajectory prediction using gan,” in Proc. Int. Conf. I-SMAC (IoT Soc. Mobile Analytics Cloud), 2020, pp. 502–507

    2020

  8. [16]

    Vehicle trajectory prediction using gen- erative adversarial network with temporal logic syntax tree features,

    X. Li, G. Rosman, I. Gilitschenski, C.-I. Vasile, J. A. DeCastro, S. Karaman, and D. Rus, “Vehicle trajectory prediction using gen- erative adversarial network with temporal logic syntax tree features,” IEEE Robot. Autom. Lett. , vol. 6, no. 2, pp. 3459–3466, 2021

    2021

  9. [18]

    Deep predictive autonomous driving using multi-agent joint trajectory prediction and traffic rules,

    K. Cho, T. Ha, G. Lee, and S. Oh, “Deep predictive autonomous driving using multi-agent joint trajectory prediction and traffic rules,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS) , 2019, pp. 2076–2081

    2019

  10. [19]

    Multi-modal anticipation of stochastic trajectories in a dynamic environment with conditional variational autoencoders,

    A. Dulian and J. C. Murray, “Multi-modal anticipation of stochastic trajectories in a dynamic environment with conditional variational autoencoders,” arXiv preprint arXiv:2103.03912 , 2021. [Online]. Available: https://arxiv.org/abs/2103.03912

    2021 arXiv

  11. [20]

    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. V eh. Technol. , vol. 65, no. 6, pp. 3981–3992, 2016

    2016

  12. [21]

    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/CVF Conf. Comput. Vis. Pattern Recog- nit. (CVPR) , 2020

    2020

  13. [22]

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

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

    2021

  14. [23]

    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 Proc. Eur . Conf. Comput. Vis. (ECCV). Berlin, Heidelberg: Springer, 2020, pp. 541–556

    2020

  15. [24]

    Lanercnn: Dis- tributed representations for graph-centric motion forecasting,

    W. Zeng, M. Liang, R. Liao, and R. Urtasun, “Lanercnn: Dis- tributed representations for graph-centric motion forecasting,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS) , 2021, pp. 532–539

    2021

  16. [25]

    Lane transformer: A high-efficiency trajectory prediction model,

    Z. Wang, J. Guo, Z. Hu, H. Zhang, J. Zhang, and J. Pu, “Lane transformer: A high-efficiency trajectory prediction model,” IEEE Open J. Intell. Transp. Syst. , vol. 4, pp. 2–13, 2023

    2023

  17. [26]

    End-to-end interpretable neural motion planner,

    W. Zeng, W. Luo, S. Suo, A. Sadat, B. Yang, S. Casas, and R. Urtasun, “End-to-end interpretable neural motion planner,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) . IEEE, Jun. 2019, pp. 8660–8669

    2019

  18. [27]

    Trajectory prediction for autonomous driving with topometric map,

    J. Xu, L. Xiao, D. Zhao, Y . Nie, and B. Dai, “Trajectory prediction for autonomous driving with topometric map,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA) . IEEE, May 2022, pp. 8403–8408

    2022

  19. [28]

    Are we ready for autonomous driving? the kitti vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in CVPR, Washington, DC, USA, 2012

    2012

  20. [29]

    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 CVPR, 2019

    2019

  21. [30]

    Multimodal trajectory predictions for autonomous driving using deep convolutional networks,

    H. Cui, V . Radosavljevic, F. C. Chou, T. H. Lin, T. Nguyen, T. K. Huang, J. Schneider, and N. Djuric, “Multimodal trajectory predictions for autonomous driving using deep convolutional networks,” in ICRA, 2018

    2018

  22. [2020]

    Available: https://arxiv.org/abs/1908.09008

    [Online]. Available: https://arxiv.org/abs/1908.09008

    1908 arXiv

Pith tools

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