Pith. sign in

REVIEW 2 major objections 4 minor 69 references

Social-Pose: Enhancing Trajectory Prediction with Human Body Pose

T0 review · 2 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 Social-Pose is a decoupled attention encoder that adds human body pose to existing trajectory predictors and consistently lowers their prediction error.

desk verdict Useful pose-encoder integration with a real but narrow capacity confound in the LSTM/GAN experiments; the Transformer result is clean and the paper deserves a careful revision. read the letter →

arxiv 2507.22742 v1 pith:6NNWHDRS submitted 2025-07-30 cs.CV

Yang Gao , Saeed Saadatnejad , Alexandre Alahi This is my paper
classification cs.CV
keywords trajectorypredictionhumanbodyposeTransformerencoderdecoupledmodulesocialinteractionautonomousdrivingrobotnavigationkeypoints
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 tries to establish that feeding a sequence of human body pose keypoints into a trajectory predictor, alongside the usual x-y positions, makes the predictor more accurate, and that this works without redesigning the underlying model. The authors propose Social-Pose, a small Transformer-based pose encoder that plugs into LSTM, GAN, MLP, and Transformer baselines with minimal modification. Across synthetic and real datasets they report consistent average and final displacement error reductions, up to 25% and 29%, and they show the pose signal helps in 2D and 3D, survives noisy or partially missing poses, and improves a downstream robot navigation simulation. A sympathetic reader should care because it suggests body language is a cheap, architecture-agnostic source of information for anticipating where people will go.

What carries the argument

The carrying mechanism is a compact Transformer pose encoder: it embeds the local (pelvis-relative) keypoint coordinates of every person at every observed time step, adds sinusoidal positional encodings to mark time, runs self-attention, and outputs one pose representation per agent. That representation is concatenated with the agent's trajectory embedding at the embedding level, rather than token-wise, so the module stays decoupled and can be dropped into different backbones. The same machinery is also what the authors use to study which joints matter, reporting that ankles, knees, wrists, and elbows carry most of the predictive signal.

What would settle it

Train the same pose-augmented model with a capacity-matched trajectory-only baseline that doubles the interaction encoder's dimension without pose input; if ADE and FDE gains disappear, the reported improvement comes from model capacity, not body keypoints.

Watch

Extended reading notes

Core claim

The central claim is that human body pose keypoints carry motion intent, especially turning, body rotation, and limb cues, that Cartesian trajectory coordinates alone miss, and that a decoupled pose encoder can extract this signal for any modern trajectory predictor. Social-Pose encodes each person's observed pelvis-relative keypoints with a Transformer, adds sinusoidal positional encodings, and concatenates the resulting pose embedding with the same person's trajectory embedding before the existing interaction encoder and decoder process the scene. The authors report consistent ADE and FDE improvements over all four baseline architecture families on the JTA, Human3.6M, and Urban datasets, with the largest gains on JTA, and show the pose-augmented Autobots outperforms the pose-based Social-Transmotion on identical input modalities. They also report that 3D poses give larger gains than 2D poses, that training with noisy poses preserves positive gains at inference, and that a social-force robot navigator using the pose-based predictor reaches its goal about 9% faster with about 37% fewer collisions.

Load-bearing premise

The observed improvements are attributed to pose content rather than to the concurrently increased model capacity: the interaction encoder's dimension is doubled after adding pose, and no capacity-matched trajectory-only baseline is reported.

Editorial extensions

If this is right

  • Pose information can improve trajectory prediction without changing the predictor's decoder or interaction model, so existing trained architectures can be upgraded by adding one encoder and retraining.
  • The computational cost of the upgrade is small: inference time for Autobots rises only about 2% while ADE and FDE improve by about 25%, making the trade-off attractive for onboard autonomous driving systems.
  • 2D pose alone still yields roughly 15% ADE and 20% FDE gains, so the method remains useful when 3D pose sensors or estimators are unavailable.
  • Training with moderately noisy poses keeps most of the benefit and prevents large collapses at inference, which matters because real pose estimators are imperfect.
  • The pose-based predictor transfers to downstream use: a robot navigator that consumes its predictions lowers collision rate by about 37% and completion time by about 9% in simulation.

Reading between the lines

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

  • Because every pose-augmented model also doubles the interaction encoder's dimension and no capacity-matched trajectory-only baseline is reported, part of the reported gain could come from extra parameters rather than pose content; a matched-capacity ablation would settle this.
  • The attention maps point to arms and legs as the most informative joints, suggesting that a sparse keypoint selection or a distilled pose embedding could retain most of the benefit at lower cost, which the authors leave as future work.
  • The gains are likely concentrated in turning and interaction-heavy maneuvers, so a per-maneuver evaluation, for example turns versus straight walking and dense versus sparse crowds, would test where pose input matters most.
  • The noisy-pose training result hints at a simple recipe: augmenting keypoints with Gaussian noise during training could become a standard robustness practice for pose-conditioned motion forecasting.
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

2 major / 4 minor

Summary. The paper proposes Social-pose, a decoupled attention-based pose encoder that is inserted into existing trajectory prediction models by concatenating pose and trajectory embeddings. The method is evaluated on LSTM-, GAN-, MLP-, and Transformer-based baselines across several datasets (JTA, Human3.6M, Urban, JRDB, ACTEV), reporting consistent ADE/FDE improvements, along with analyses of 2D vs. 3D pose, noisy and partial pose, joint attention, and a robot navigation application. The central claim is that pose information, when encoded with the proposed module, yields architecture-agnostic accuracy gains.

Significance. If the stated improvements were robust and properly controlled, the paper would provide a useful, cheap plug-in for trajectory prediction. The experiments are extensive, covering several architectures and datasets, and the ablation on 2D vs. 3D pose and the robot navigation study are valuable additions. However, the primary evidence for the central claim is weakened by a capacity confound in the LSTM/GAN comparisons and by the absence of variance reporting. The work is therefore of potential interest to the trajectory prediction community, but the contribution is not yet convincingly established.

major comments (2)
  1. [Section IV-D, Table I] The central claim that pose information consistently improves ADE/FDE is confounded by a capacity increase. The paper states, "After incorporating pose information, we double the interaction encoder's dimension to enable the model to capture both trajectory and pose interactions." Thus for Vanilla-LSTM, Social-LSTM, Directional-LSTM, Dir-social-LSTM, Social-GAN, and Directional-GAN, the pose-augmented models use a 512-dimensional interaction encoder while the trajectory-only baselines use 256 dimensions. Since no capacity-matched trajectory-only baseline is reported, the observed improvements (e.g., Social-LSTM 1.21→1.11 ADE; Social-GAN 1.66→1.49 ADE on JTA) could be due entirely or in part to the extra parameters rather than to body pose. The sentence in the same subsection claiming "we ensure a fair comparison" is therefore not supported. Please add a trajectory-only baseline with the same 512-dimensional interaction encoder (or otherwise match capacity) to isolate the effect of pose content.
  2. [Section IV-E, Tables I and III-VI] All quantitative results are reported as single-point estimates. On Human3.6M and the Urban dataset, the differences are very small (e.g., EqMotion on Human3.6M: 0.51/0.81 vs. 0.48/0.79; Social-LSTM on Urban: 0.58/1.06 vs. 0.57/1.04). Without standard deviations across multiple seeds or significance tests, these differences cannot be distinguished from training noise. This is especially important for the claim that pose helps "all four types of architectures." Please report mean and standard deviation over at least three random seeds, or provide significance tests, for the main comparisons.
minor comments (4)
  1. [Table II] The unit "miliseconds" is misspelled; it should be "milliseconds."
  2. [Section IV-G-6, Table VII] The phrase "we further explore whether a cross-attention module ... outperforms direct concatenation" reads as an open question, but the results show that direct concatenation performs better. Please rephrase to avoid ambiguity, e.g., "we compare a cross-attention module against direct concatenation."
  3. [Figure 6] The vertical axis is not labeled; please add a label such as "Attention weight (%)" to make the figure self-contained.
  4. [Section IV-E] The sentence "with gains of up to 25% and 29%, respectively" refers specifically to the Autobots + 3D Pose model, but the preceding text may lead readers to believe these maximum gains apply to all baselines. Please clarify which model achieves these gains.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: pose-augmented gains are measured against external datasets; the capacity confound is an experimental-control issue, not a circular derivation.

full rationale

The paper's central claim is empirical: adding body-pose keypoints through the Social-pose encoder improves ADE/FDE on external datasets. The derivation chain is not circular. The pose encoder is defined independently (Eqs. 1-4: sinusoidal positional encoding, transformer encoder, concatenation of H_pose,i with H_traj,i), and its value is established by training pose-augmented versions of LSTM, GAN, MLP, and Transformer baselines and comparing them with the corresponding trajectory-only versions on JTA, Human3.6M, Urban, JRDB, and ACTEV. No evaluation metric is used as a training target or fitting objective in a way that would make the reported gains true by construction; no fitted parameter is renamed as a prediction; no uniqueness theorem or prior claim by the same authors is invoked to exclude alternative explanations. The many self-citations (Social-LSTM, Social-GAN, Trajnet++, Social-Transmotion, CrowdNav, Multi-Transmotion) function as baselines, tools, or related work, not as load-bearing premises: for example, the claim to beat Social-Transmotion is supported by Table I numbers produced in this paper, not by the citation itself. The most serious validity concern is that Section IV-D doubles the interaction-encoder dimension for pose-augmented LSTM/GAN models, so part of the gain may reflect added capacity rather than pose content; however, that is a controlled-variable and attribution problem, not circularity, because the pose model is still a distinct model evaluated on held-out data. Under the rule that circularity requires exhibiting a reduction of a claimed result to its inputs by construction or by a self-citation chain, this paper is not circular.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on experimental evidence rather than derivation. The main counted free parameters are design choices that affect attribution of the gains. No new physical entities are introduced; the module is an architectural invention.

free parameters (2)
  • interaction encoder dimension scale = 2x (doubled when pose is added)
    The paper doubles the interaction encoder's pooling dimension from 256 to 512 for all pose-augmented models (Section IV-D). This is a hand-chosen design change that confounds the effect of pose with increased model capacity.
  • pose noise standard deviation (training) = 0.1
    In the noisy-pose robustness experiment (Section IV-G.3), models are trained with Gaussian noise (mean 0, std=0.1) added to 50% of scenes; this value is chosen by hand and affects reported robustness.
assumptions (3)
  • domain assumption Ground-truth 3D poses are available as input at test time.
    The main experiments use ground-truth 3D pose keypoints from JTA, Human3.6M, and Urban datasets; the paper does not evaluate with poses from an actual detector, only with synthetic noise.
  • domain assumption The Trajnet++ toolbox preprocessing and scene categorization are appropriate for all datasets.
    Datasets are rebalanced by trajectory type following the Trajnet++ benchmark; differences in original annotation formats may introduce bias.
  • domain assumption Transformer self-attention can represent pose dynamics sufficient for trajectory prediction.
    The method assumes the pose encoder's attention mechanism captures motion-relevant cues; the paper provides an attention-map analysis but no formal guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Social-Pose: Enhancing Trajectory Prediction with Human Body Pose." pith.science (2026). https://pith.science/paper/6NNWHDRS

@misc{pith2026250722742,
  author       = {Pith},
  title        = {Pith review of: Social-Pose: Enhancing Trajectory Prediction with Human Body Pose},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NNWHDRS}},
  note         = {Machine review of arXiv:2507.22742}
}
read the original abstract

Accurate human trajectory prediction is one of the most crucial tasks for autonomous driving, ensuring its safety. Yet, existing models often fail to fully leverage the visual cues that humans subconsciously communicate when navigating the space. In this work, we study the benefits of predicting human trajectories using human body poses instead of solely their Cartesian space locations in time. We propose `Social-pose', an attention-based pose encoder that effectively captures the poses of all humans in a scene and their social relations. Our method can be integrated into various trajectory prediction architectures. We have conducted extensive experiments on state-of-the-art models (based on LSTM, GAN, MLP, and Transformer), and showed improvements over all of them on synthetic (Joint Track Auto) and real (Human3.6M, Pedestrians and Cyclists in Road Traffic, and JRDB) datasets. We also explored the advantages of using 2D versus 3D poses, as well as the effect of noisy poses and the application of our pose-based predictor in robot navigation scenarios.

Figures

Figures reproduced from arXiv: 2507.22742 by the authors.

Figure 1
Figure 1. Given observed trajectories and pose keypoints of all [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Social-pose: Our human pose encoder enhances trajectory prediction. It takes the sequence of observed poses of all people in the scene and generates a rich representation. This enriched information aids the trajectory decoder in predicting more accurate trajectories. learns a representation of pose cues and integrates it with the standard trajectory-encoded representation. The integration of this module requires a s… view at source ↗
Figure 3
Figure 3. Comparison between work [45] and our Social-Pose. [45] (a) fuses embeddings in token-wise concatenation. Our Social-Pose (b) uses embedding-wise concatenation for better compatibility with different trajectory predictors. connection. Subsequently, the features go through a feed￾forward layer, and once again, a residual connection is ap￾plied. Then, the transformer encoder outputs the latent pose representations: Hpo… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative examples on the JTA [13] dataset. Each example depicts pedestrian trajectories within a specific scene. For the primary agent, the ground truth is shown in green, the baseline model’s prediction in red, and the pose-based model’s prediction in blue. All oth…
Figure 5
Figure 5. Figure 5: Qualitative examples on the Human3.6M [26] dataset. For the primary agent, the ground truth is shown in green, the model’s prediction in red, and the pose-based model’s prediction in blue. All other agents are depicted in gray. lows us to identify which frames and pose…
Figure 6
Figure 6. Figure 6: Attention map for joints. Spatially, the arms and legs are more significant in trajectory prediction. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results of the robot navigation task without [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 65 canonical work pages

  1. [1]

    Social lstm: Human trajec- tory prediction in crowded spaces

    Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Ro- bicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajec- tory prediction in crowded spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) , pages 961–971, 2016. 1, 2, 5, 6

    work page 2016

  2. [2]

    Inject- ing knowledge in data-driven vehicle trajectory predictors

    Mohammadhossein Bahari, Ismail Nejjar, and Alexandre Alahi. Inject- ing knowledge in data-driven vehicle trajectory predictors. Transporta- tion Research Part C: Emerging Technologies , 128:103010, 2021. 1

    work page 2021

  3. [3]

    Bayesian intention inference for trajectory prediction with an unknown goal destination

    Graeme Best and Robert Fitch. Bayesian intention inference for trajectory prediction with an unknown goal destination. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 5817–5823. IEEE, 2015. 2

    work page 2015

  4. [4]

    Mpc-pf: socially and spatially aware object trajectory prediction for autonomous driving systems using potential fields

    Neel P Bhatt, Amir Khajepour, and Ehsan Hashemi. Mpc-pf: socially and spatially aware object trajectory prediction for autonomous driving systems using potential fields. IEEE Transactions on Intelligent Trans- portation Systems, 24(5):5351–5361, 2023. 2

    work page 2023

  5. [5]

    Euro-pvi: Pedestrian vehicle interactions in dense urban centers

    Apratim Bhattacharyya, Daniel Olmeda Reino, Mario Fritz, and Bernt Schiele. Euro-pvi: Pedestrian vehicle interactions in dense urban centers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6408–6417, 2021. 2

    work page 2021

  6. [6]

    Pedes- trian intention prediction: A multi-task perspective

    Smail Ait Bouhsain, Saeed Saadatnejad, and Alexandre Alahi. Pedes- trian intention prediction: A multi-task perspective. arXiv preprint arXiv:2010.10270, 2020. 2

    arXiv 2010

  7. [7]

    Yves V on Cramon, and Ricarda Schubotz

    Andreja Bubic, D. Yves V on Cramon, and Ricarda Schubotz. Prediction, cognition and the brain. Frontiers in Human Neuroscience, 4:25, 2010. 1

    work page 2010

  8. [8]

    Crowd- robot interaction: Crowd-aware robot navigation with attention-based deep reinforcement learning

    Changan Chen, Yuejiang Liu, Sven Kreiss, and Alexandre Alahi. Crowd- robot interaction: Crowd-aware robot navigation with attention-based deep reinforcement learning. In 2019 International Conference on Robotics and Automation (ICRA) , pages 6015–6022. IEEE, 2019. 1, 9

    work page 2019

Show all 69 references
  1. [9]

    Vulnerable road user trajectory prediction for autonomous driving using a data-driven integrated approach

    Hao Chen, Yinhua Liu, Chuan Hu, and Xi Zhang. Vulnerable road user trajectory prediction for autonomous driving using a data-driven integrated approach. IEEE Transactions on Intelligent Transportation Systems, 24(7):7306–7317, 2023. 2

    2023

  2. [10]

    Pedestrian trajectory predic- tion in heterogeneous traffic using pose keypoints-based convolutional encoder-decoder network

    Kai Chen, Xiao Song, and Xiaoxiang Ren. Pedestrian trajectory predic- tion in heterogeneous traffic using pose keypoints-based convolutional encoder-decoder network. IEEE Transactions on Circuits and Systems for Video Technology, 31(5):1764–1775, 2020. 2

    2020

  3. [11]

    Stochastic non-autoregressive transformer-based multi-modal pedestrian trajectory prediction for intelligent vehicles

    Xiaobo Chen, Huanjia Zhang, Fuwen Deng, Jun Liang, and Jian Yang. Stochastic non-autoregressive transformer-based multi-modal pedestrian trajectory prediction for intelligent vehicles. IEEE Transactions on Intelligent Transportation Systems, 2023. 2

    2023

  4. [12]

    Mg-gan: A multi- generator model preventing out-of-distribution samples in pedestrian trajectory prediction

    Patrick Dendorfer, Sven Elflein, and Laura Leal-Taixé. Mg-gan: A multi- generator model preventing out-of-distribution samples in pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 13158–13167, 2021. 1

    2021

  5. [13]

    Learning to detect and track visible and occluded body joints in a virtual world

    Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In Proceedings of the European conference on computer vision (ECCV) , pages 430–446, 2018. 2, 4,...

    2018

  6. [14]

    Summary and reflections on pedestrian trajectory prediction in the field of autonomous driving

    Zheng Fu, Kun Jiang, Chuchu Xie, Yuhang Xu, Jin Huang, and Diange Yang. Summary and reflections on pedestrian trajectory prediction in the field of autonomous driving. IEEE Transactions on Intelligent Vehicles,

  7. [15]

    Multi-transmotion: Pre-trained model for human motion prediction

    Yang Gao, Po-Chien Luan, and Alexandre Alahi. Multi-transmotion: Pre-trained model for human motion prediction. In 8th Annual Confer- ence on Robot Learning , 2024. 1, 2

    2024

  8. [16]

    Adaptive and simultaneous trajectory prediction for heterogeneous agents via transferable hierarchical transformer network

    Maosi Geng, Junyi Li, Chuangjia Li, Ningke Xie, Xiqun Chen, and Der-Horng Lee. Adaptive and simultaneous trajectory prediction for heterogeneous agents via transferable hierarchical transformer network. IEEE Transactions on Intelligent Transportation Systems, 24(10):11479– 114...

    2023

  9. [17]

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

    Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Al- don D’Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In International Conference on Learning Representati...

    2022

  10. [18]

    Transformer networks for trajectory forecasting

    Francesco Giuliari, Irtiza Hasan, Marco Cristani, and Fabio Galasso. Transformer networks for trajectory forecasting. In International con- ference on pattern recognition (ICPR), pages 10335–10342. IEEE, 2021. 2, 5, 6

    2021

  11. [19]

    Stochastic trajectory prediction via motion indeterminacy diffusion

    Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 17113– 17122, June 2022. 2

    2022

  12. [20]

    Social gan: Socially acceptable trajectories with generative adversarial networks

    Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) , pages 2255–2264,

  13. [21]

    Forecasting people trajectories and head poses by jointly reasoning on tracklets and vislets

    Irtiza Hasan, Francesco Setti, Theodore Tsesmelis, Vasileios Bela- giannis, Sikandar Amin, Alessio Del Bue, Marco Cristani, and Fabio Galasso. Forecasting people trajectories and head poses by jointly reasoning on tracklets and vislets. IEEE transactions on pattern analysis an...

    2019

  14. [22]

    Social force model for pedestrian dynamics

    Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E , 51(5):4282, 1995. 1, 2, 9

    1995

  15. [23]

    Collaborative motion prediction via neural motion message passing

    Yue Hu, Siheng Chen, Ya Zhang, and Xiao Gu. Collaborative motion prediction via neural motion message passing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6319–6328, 2020. 2

    2020

  16. [24]

    Fully decoupling trajectory and scene encoding for lightweight heatmap- oriented trajectory prediction

    Renhao Huang, Jingtao Ding, Maurice Pagnucco, and Yang Song. Fully decoupling trajectory and scene encoding for lightweight heatmap- oriented trajectory prediction. IEEE Robotics and Automation Letters ,

  17. [25]

    Stgat: Modeling spatial-temporal interactions for human trajectory pre- diction

    Yingfan Huang, Huikun Bi, Zhaoxin Li, Tianlu Mao, and Zhaoqi Wang. Stgat: Modeling spatial-temporal interactions for human trajectory pre- diction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages 6272–6281, 2019. 2

    2019

  18. [26]

    Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments

    Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence , 36(7):1325–1339, jul 2014. 2, 4, 5, 6, 7

    2014

  19. [27]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 , 2014. 5

    2014 arXiv

  20. [28]

    Human trajectory forecasting in crowds: A deep learning perspective

    Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human trajectory forecasting in crowds: A deep learning perspective. IEEE Transactions on Intelligent Transportation Systems , 2021. 2, 4, 5, 6

    2021

  21. [29]

    PifPaf: Composite fields for human pose estimation

    Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. PifPaf: Composite fields for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2

    2019

  22. [30]

    Pose and semantic map based probabilistic forecast of vulnerable road users trajectories

    Viktor Kress, Fabian Jeske, Stefan Zernetsch, Konrad Doll, and Bernhard Sick. Pose and semantic map based probabilistic forecast of vulnerable road users trajectories. IEEE Transactions on Intelligent Vehicles, 2022. JOURNAL OF LATEX CLASS FILES 11 1, 2, 4, 5, 6, 9

    2022

  23. [31]

    Hierarchical light transformer ensembles for multimodal trajectory fore- casting

    Adrien Lafage, Mathieu Barbier, Gianni Franchi, and David Filliat. Hierarchical light transformer ensembles for multimodal trajectory fore- casting. arXiv preprint arXiv:2403.17678 , 2024. 2

    2024 arXiv

  24. [32]

    Traj-llm: A new exploration for empowering trajectory prediction with pre-trained large language models

    Zhengxing Lan, Lingshan Liu, Bo Fan, Yisheng Lv, Yilong Ren, and Zhiyong Cui. Traj-llm: A new exploration for empowering trajectory prediction with pre-trained large language models. IEEE Transactions on Intelligent Vehicles, 2024. 2

    2024

  25. [33]

    Graph-based spatial transformer with memory replay for multi-future pedestrian trajectory prediction

    Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multi-future pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 2231–2241, 2022. 2

    2022

  26. [34]

    A pedestrian trajectory prediction model for right-turn unsignalized intersections based on game theory

    Wenli Li, Yinan Zhang, Lingxi Li, Yisheng Lv, and Mengxin Wang. A pedestrian trajectory prediction model for right-turn unsignalized intersections based on game theory. IEEE Transactions on Intelligent Transportation Systems, 2024. 2

    2024

  27. [35]

    Peeking into the future: Predicting future person activities and locations in videos

    Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5725–573...

    2019

  28. [36]

    Egocentric vulnerable road users trajectory prediction with incomplete observation

    Hui Liu, Chunsheng Liu, Faliang Chang, Yansha Lu, and Minhang Liu. Egocentric vulnerable road users trajectory prediction with incomplete observation. IEEE Transactions on Intelligent Transportation Systems ,

  29. [37]

    Unified human localization and trajectory prediction with monocular vision

    Po-Chien Luan, Yang Gao, Céline Demonsant, and Alexandre Alahi. Unified human localization and trajectory prediction with monocular vision. arXiv preprint arXiv:2503.03535 , 2025. 2

    2025 arXiv

  30. [38]

    A simple yet effective baseline for 3d human pose estimation

    Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2640–2649, 2017. 4

    2017

  31. [39]

    Dag-net: Double attentive graph neural network for trajectory forecast- ing

    Alessio Monti, Alessia Bertugli, Simone Calderara, and Rita Cucchiara. Dag-net: Double attentive graph neural network for trajectory forecast- ing. In International Conference on Pattern Recognition (ICPR) , pages 2551–2558. IEEE, 2021. 2

    2021

  32. [40]

    A large-scale benchmark dataset for event recognition in surveillance video

    Sangmin Oh, Anthony Hoogs, Amitha Perera, Naresh Cuntoor, Chia- Chih Chen, Jong Taek Lee, Saurajit Mukherjee, Jake K Aggarwal, Hyungtae Lee, Larry Davis, et al. A large-scale benchmark dataset for event recognition in surveillance video. In CVPR 2011 , pages 3153–

    2011

  33. [41]

    Sim-to-real causal transfer: A metric learning approach to causally-aware interaction representations

    Ahmad Rahimi, Po-Chien Luan, Yuejiang Liu, Frano Raji ˇc, and Alexan- dre Alahi. Sim-to-real causal transfer: A metric learning approach to causally-aware interaction representations. In Proceedings of the Computer Vision and Pattern Recognition Conference , pages 17271– 17281...

    2025

  34. [42]

    Pedestrian intention prediction: A convolutional bottom-up multi-task approach

    Haziq Razali, Taylor Mordan, and Alexandre Alahi. Pedestrian intention prediction: A convolutional bottom-up multi-task approach. Transporta- tion research part C: emerging technologies , 130:103259, 2021. 2

    2021

  35. [43]

    Tessetrack: End-to-end learnable multi- person articulated 3d pose tracking

    N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G Narasimhan. Tessetrack: End-to-end learnable multi- person articulated 3d pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15190–...

    2021

  36. [44]

    Are socially-aware trajectory prediction models really socially-aware? Transportation research part C: emerging technologies, 141:103705, 2022

    Saeed Saadatnejad, Mohammadhossein Bahari, Pedram Khorsandi, Mohammad Saneian, Seyed-Mohsen Moosavi-Dezfooli, and Alexan- dre Alahi. Are socially-aware trajectory prediction models really socially-aware? Transportation research part C: emerging technologies, 141:103705, 2022. 1

    2022

  37. [45]

    Social-transmotion: Promptable human trajectory prediction

    Saeed Saadatnejad, Yang Gao, Kaouther Messaoud, and Alexandre Alahi. Social-transmotion: Promptable human trajectory prediction. In International Conference on Learning Representations (ICLR) , 2024. 1, 2, 4, 5, 6

    2024

  38. [46]

    Pedestrian 3d bounding box prediction

    Saeed Saadatnejad, Yi Zhou Ju, and Alexandre Alahi. Pedestrian 3d bounding box prediction. arXiv preprint arXiv:2206.14195 , 2022. 2

    2022 arXiv

  39. [47]

    Toward reliable human pose forecasting with uncertainty

    Saeed Saadatnejad, Mehrshad Mirmohammadi, Matin Daghyani, Parham Saremi, Yashar Zoroofchi Benisi, Amirhossein Alimohammadi, Zahra Tehraninasab, Taylor Mordan, and Alexandre Alahi. Toward reliable human pose forecasting with uncertainty. IEEE Robotics and Automation Letters, 2024. 2

    2024

  40. [48]

    Sophie: An attentive gan for predicting paths compliant to social and physical constraints

    Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognit...

    2019

  41. [49]

    Intent pre- diction of pedestrians via motion trajectories using stacked recurrent neural networks

    Khaled Saleh, Mohammed Hossny, and Saeid Nahavandi. Intent pre- diction of pedestrians via motion trajectories using stacked recurrent neural networks. IEEE Transactions on Intelligent Vehicles , 3(4):414– 424, 2018. 2

    2018

  42. [50]

    Human trajectory prediction with momentary observation

    Jianhua Sun, Yuxuan Li, Liang Chai, Hao-Shu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with momentary observation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6467–6476, 2022. 2

    2022

  43. [51]

    Unfreezing the robot: Navigation in dense, interacting crowds

    Peter Trautman and Andreas Krause. Unfreezing the robot: Navigation in dense, interacting crowds. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 797–803. IEEE, 2010. 1

    2010

  44. [52]

    Evaluating pedes- trian trajectory prediction methods with respect to autonomous driving

    Nico Uhlemann, Felix Fent, and Markus Lienkamp. Evaluating pedes- trian trajectory prediction methods with respect to autonomous driving. IEEE Transactions on Intelligent Transportation Systems , 2024. 2

    2024

  45. [53]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30,

  46. [54]

    Jrdb-pose: A large-scale dataset for multi-person pose estimation and tracking

    Edward Vendrow, Duy Tho Le, Jianfei Cai, and Hamid Rezatofighi. Jrdb-pose: A large-scale dataset for multi-person pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4811–4820, 2023. 4, 8, 9

    2023

  47. [55]

    Canonpose: Self-supervised monocular 3d human pose estimation in the wild

    Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13294–13304,

  48. [56]

    Mtp- go: Graph-based probabilistic multi-agent trajectory prediction with neural odes

    Theodor Westny, Joel Oskarsson, Björn Olofsson, and Erik Frisk. Mtp- go: Graph-based probabilistic multi-agent trajectory prediction with neural odes. IEEE Transactions on Intelligent Vehicles, 8(9):4223–4236,

  49. [57]

    Msn: multi-style network for trajectory prediction

    Conghao Wong, Beihao Xia, Qinmu Peng, Wei Yuan, and Xinge You. Msn: multi-style network for trajectory prediction. IEEE Transactions on Intelligent Transportation Systems , 24(9):9751–9766, 2023. 2

    2023

  50. [58]

    Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning

    Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page...

    2023

  51. [59]

    Future person localization in first-person videos

    Takuma Yagi, Karttikeya Mangalam, Ryo Yonetani, and Yoichi Sato. Future person localization in first-person videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 7593–7602, 2018. 2

    2018

  52. [60]

    A multi-task learning network with a collision-aware graph transformer for traffic-agents trajectory prediction

    Biao Yang, Fucheng Fan, Rongrong Ni, Hai Wang, Ammar Jafaripourn- imchahi, and Hongyu Hu. A multi-task learning network with a collision-aware graph transformer for traffic-agents trajectory prediction. IEEE Transactions on Intelligent Transportation Systems , 2024. 2

    2024

  53. [61]

    Long-short term spatio-temporal aggregation for trajectory prediction

    Cuiliu Yang and Zhao Pei. Long-short term spatio-temporal aggregation for trajectory prediction. IEEE Transactions on Intelligent Transporta- tion Systems, 24(4):4114–4126, 2023. 2

    2023

  54. [62]

    Spatio- temporal graph transformer networks for pedestrian trajectory predic- tion

    Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio- temporal graph transformer networks for pedestrian trajectory predic- tion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 507–523, 2020. 2

    2020

  55. [63]

    Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M. Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9813–9823, 2021. 2

    2021

  56. [64]

    Learning the pedestrian-vehicle inter- action for pedestrian trajectory prediction

    Chi Zhang and Christian Berger. Learning the pedestrian-vehicle inter- action for pedestrian trajectory prediction. In International Conference on Control, Automation and Robotics (ICCAR) , pages 230–236. IEEE,

  57. [65]

    Sr-lstm: State refinement for lstm towards pedestrian trajectory prediction

    Pu Zhang, Wanli Ouyang, Pengfei Zhang, Jianru Xue, and Nanning Zheng. Sr-lstm: State refinement for lstm towards pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 12085–12094, 2019. 2

    2019

  58. [66]

    Bip- tree: Tree variant with behavioral intention perception for heterogeneous trajectory prediction

    Yuzhen Zhang, Weizhi Guo, Junning Su, Pei Lv, and Mingliang Xu. Bip- tree: Tree variant with behavioral intention perception for heterogeneous trajectory prediction. IEEE Transactions on Intelligent Transportation Systems, 24(9):9584–9598, 2023. 2

    2023

  59. [67]

    Trajgat: A map-embedded graph attention network for real-time vehicle trajectory imputation of roadside perception

    Cong Zhao, Andi Song, Yuchuan Du, and Biao Yang. Trajgat: A map-embedded graph attention network for real-time vehicle trajectory imputation of roadside perception. Transportation Research Part C: Emerging Technologies, 142:103787, 2022. 1

    2022

  60. [68]

    Visual exposes you: pedestrian trajectory prediction meets visual intention

    Xian Zhong, Xu Yan, Zhengwei Yang, Wenxin Huang, Kui Jiang, Ryan Wen Liu, and Zheng Wang. Visual exposes you: pedestrian trajectory prediction meets visual intention. IEEE Transactions on Intelligent Transportation Systems, 24(9):9390–9400, 2023. 2

    2023

  61. [69]

    Trajpred: Trajectory prediction with region-based relation learning.IEEE Transactions on Intelligent Transportation Systems , 2024

    Chen Zhou, Ghassan AlRegib, Armin Parchami, and Kunjan Singh. Trajpred: Trajectory prediction with region-based relation learning.IEEE Transactions on Intelligent Transportation Systems , 2024. 2 12 JOURNAL OF LATEX CLASS FILES Yang Gao received the B.E. degree from Shanghai J...

    2024

Pith tools

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