Pith. sign in

REVIEW 5 major objections 6 minor 36 references

Pedestrian Intention Prediction via Vision-Language Foundation Models

T0 review · 5 major / 6 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 The paper claims that prompting vision-language models with ego-vehicle speed and its time variation significantly improves pedestrian crossing-intention prediction, with gains up to 19.8% from context and a further 12.5% from automatic…

desk verdict A plausible prompt-engineering recipe for VLFMs in pedestrian intention prediction, but the headline accuracy claims are validation-set artifacts and the vision-baseline comparison is overstated. read the letter →

arxiv 2507.04141 v1 pith:Z5ZPKYUX submitted 2025-07-05 cs.CV cs.AIcs.ETcs.LGcs.RO

Mohsen Azarmi , Mahdi Rezaei , He Wang This is my paper
classification cs.CVcs.AIcs.ETcs.LGcs.RO
keywords pedestrianintentionpredictionvision-languagefoundationmodelspromptengineeringautonomousdrivingvehicledynamicsGPT-4VJAADdatasetPIE
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

The paper sets out to show that vision-language foundation models (VLFMs) can predict pedestrian crossing intention more accurately than conventional vision-only models when their prompts are built hierarchically and enriched with ego-vehicle speed, especially speed changes over time. On three public datasets the authors report that adding such context improves accuracy by up to 19.8%, and that automatically optimized prompts add a further 12.5%. The work matters because if true it gives autonomous vehicles a way to use the contextual reasoning of large multimodal models rather than relying solely on learned visual features.

What carries the argument

The key device is the hierarchical prompt template: a sequence of prompt types that first establishes the model's role as an autonomous vehicle observing a front-view camera, then directs it to watch the red-boxed pedestrian's posture and body orientation, and then feeds in ego-vehicle dynamics in one of three forms—numeric speed, descriptive speed, or time-conscious descriptive speed describing the change over an interval. These templates are scored and refined by the Automatic Prompt Engineer (APE), whose score combines execution accuracy with the log probability of the correct label, yielding an optimized prompt $\rho^*$. The visual input is a clip of 16 frames at 30 fps with timestamps and a red bounding box around the target pedestrian.

What would settle it

Re-run the vision baselines (for example PCPA, GraphPlus, and PIP-Net) on the exact same test frames, splits, and metrics used for the VLFMs; if the 0.70-versus-0.74 JAAD accuracy gap shrinks to noise when the protocol is matched, the claimed superiority collapses. With only 126 JAAD test samples, a shift of about five correct predictions would flip the ranking.

Watch

Extended reading notes

Core claim

The paper's central claim is that a hierarchy of prompts—role, physical cues, and vehicle dynamics—progressively improves a VLFM's crossing-intention judgments, and that the richest prompt type (time-conscious descriptive speed) works best. On validation, the role-only prompt scores 0.54 accuracy; adding physical cues reaches 0.62, numeric speed 0.64, descriptive speed 0.70, and time-conscious speed 0.72. With optimized prompts, the largest model (GPT-4V) reaches 0.74 accuracy on JAAD and FU-PIP and 0.81 on PIE, surpassing the previous VLM approaches and most vision baselines in the comparison table, with PIP-Net retaining the highest PIE accuracy.

Load-bearing premise

The headline comparison assumes the vision-model scores quoted from earlier papers were produced under the same evaluation protocol as the new VLM results, yet the paper does not re-run those baselines and reports no significance tests.

Editorial extensions

If this is right

  • Vehicle-speed context, especially time-conscious phrasing, becomes a reusable prompt ingredient for VLM-based intention prediction, improving validation accuracy from 0.62 (physical cues only) to 0.72.
  • Automatic prompt engineering transfers across VLM architectures: extending APE to smaller models for extra iterations did not improve their top prompts, so the optimized templates are not GPT-4V-specific.
  • Larger VLFMs scale with contextual prompting, with GPT-4V outperforming GPT-4 mini and LLaVA-Next on all three datasets.
  • VLFMs show their largest edge over vision baselines on the FU-PIP intersection dataset, where context reasoning matters most, while a strong vision baseline (PIP-Net) remains competitive on PIE.

Reading between the lines

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

  • A clean ablation removing the red bounding box and timestamps from the input frames would test how much of the gain comes from the visual annotation instead of the language context; the paper always presents both together.
  • Because the quoted gains are relative accuracy improvements on small test sets, the practical effect at the level of individual JAAD or FU-PIP clips is a handful of correctly flipped predictions; deployment decisions would need larger-scale evaluation.
  • The authors position the result as VLFMs beating vision models, but the paper does not account for compute cost or latency; a fair comparison for a vehicle would weigh accuracy against the feasibility of running a 1.8T-parameter model onboard.
  • A testable extension would be distilling the optimized prompts and the reasoning they elicit into a smaller on-device model, since the paper already shows 3B and 7B models benefit from the same prompts, just with lower accuracy.
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

5 major / 6 minor

Summary. The paper adapts four vision-language foundation models (GPT-4V, GPT-4 mini, LLaVA-Next 7B and 3B) to pedestrian crossing-intention prediction. It introduces hierarchical prompt templates that progressively add physical cues (PB), numeric speed (PDs), descriptive speed (PDd), and time-conscious speed dynamics (PDt), and then optimizes these templates with the Automatic Prompt Engineer (APE) framework. Experiments on JAAD, PIE, and FU-PIP report prompt-hierarchy improvements on validation sets up to 0.72 accuracy and compare optimized VLFMs against vision-based baselines on test sets. The central claims are that temporal vehicle-dynamics prompts improve accuracy up to 19.8% and that VLFMs outperform conventional vision-based models.

Significance. Systematic prompt engineering for VLFMs in pedestrian intention prediction is a timely and practically relevant problem, and the paper's hierarchical template design is a reasonable contribution. The comparison across four VLFMs and three datasets is broad, and the identification of high- and low-performing prompt keywords is useful engineering knowledge. However, the paper's headline quantitative claims are currently not supported by the evidence as presented: the prompt-hierarchy improvement is measured on the validation set used for prompt selection, and the test-set comparison actually contradicts the claim of consistent VLM superiority (PIP-Net beats GPT-4V on PIE, and GraphPlus has higher F1 on JAAD). The paper would be publishable after a substantial revision that grounds the claims in test-set results and corrected comparisons.

major comments (5)
  1. [III.B, Fig. 2, Table I] The central claim of 'up to 19.8%' accuracy gains from hierarchical prompts and vehicle-speed context is based on validation-set results, not held-out test frames. Section III.B states that APE optimizes on the validation sets, Fig. 2 shows validation performance, and Table I, which contains PR 0.54 to PDt 0.72, is presented in the same subsection without any statement that it uses test frames. Section III.D separately introduces test sets for Table II, using only the optimized prompts. Thus the quantitative prompt-ablation claim is subject to selection bias and has no independent test-set support. The authors must either report test-set results for each prompt template (PR through PDt) or explicitly restrict the claim to validation performance.
  2. [Table II, Section III.D] The statement that VLFMs 'consistently outperform' conventional vision-based models is contradicted by the paper's own numbers. PIP-Net achieves 0.91 Acc and 0.84 F1 on PIE versus GPT-4V's 0.81 Acc and 0.73 F1, and GraphPlus achieves 0.76 F1 on JAAD versus GPT-4V's 0.66 F1. The comparison is also not controlled: the vision-baseline numbers are taken from prior work without re-running them under the same evaluation protocol, and no confidence intervals or significance tests are reported, despite test sets as small as 94 samples (FU-PIP) and 126 samples (JAAD). The superiority claim must be rewritten to match the actual table and supported by uncertainty quantification or re-evaluation.
  3. [III.B, Table I, Section III.A] There is an internal inconsistency about JAAD speed annotations. Section III.B says that 'due to the limitations of the JAAD dataset, which only provides descriptive speed information, the experiments on this dataset are restricted to descriptive speed templates (PDd)'. Yet Table I reports results for PDs and PDt, and M=365 equals the sum of the JAAD (32), PIE (243), and FU-PIP (90) validation samples. As written, it is unclear whether Table I aggregates across all three datasets despite the JAAD restriction, or whether PDs/PDt were evaluated only on PIE and FU-PIP with a different M. This provenance issue must be clarified.
  4. [II.C, Eq. (1)] The prompt-scoring function fscore(ρj) depends on a weighting parameter α, but α is never specified in the paper. Since this score determines which prompts are selected in APE, and all downstream results depend on that selection, the unreported α is a load-bearing hyperparameter. The authors should report its value and ideally show robustness of the main results to α choices.
  5. [Abstract vs. Conclusion] The abstract says APE yielded '12.5% further accuracy gains', but the conclusion says APE achieved 'an additional 5.3% gain', and the same sentence also mentions 20 additional iterations for other VLFMs while Section III.B says 10 iterations. These numerical and procedural inconsistencies need reconciliation; otherwise the reader cannot tell which number is the actual reported result.
minor comments (6)
  1. [Table II] The model name 'StakedRNN' is a typo for 'StackedRNN'.
  2. [III.C.1] The phrase 'which over presents the designated intersection scenarios' should be 'which over-represents the designated intersection scenarios'.
  3. [Throughout] Spelling and naming are inconsistent: 'LLaV A-Next' should be 'LLaVA-Next', and 'GraphPlus' in Table II and text is referred to as 'Pedestrian Graph+' in reference [8].
  4. [II.A.2] The 'time interval' placeholder in the PDt template is never given a concrete value; the paper should state what interval was used (e.g., the 16-frame observation window) and whether it varied.
  5. [III.B] The statement that M=365 is the number of samples should be made explicit as the sum of the three validation sets, and the paper should say whether APE was run on the pooled validation set or separately per dataset.
  6. [II.C] The references for model parameter counts (GPT-4V 1.8T, GPT-4 mini 8B, LLaVA-Next 7B) are not provided in the cited works; please add sources for these specific sizes.

Circularity Check

1 steps flagged · score 6.0 of 10

Prompt-gain claims reduce to APE's validation objective; final VLM comparison is independent.

  1. fitted input called prediction [Sec. II.C (APE scoring) and Sec. III.B / Table I]
    "fscore(ρj) = α fexec(ρj) + (1− α) flogprob(ρj). ρ∗ = arg max ρj fscore(ρj). The optimisation process is conducted with M = 365 ... across the validation sets of all three datasets. Table I represents the evaluation of the best-performing prompt in each template pool ... The PDt template ... performs best with an accuracy of 0.72 ... In contrast, the PR template ... performs the worst with an accuracy of 0.54."

    The 0.54→0.72 accuracy improvement advertised in the abstract and conclusion is the fexec component of the fscore that APE maximized when it selected ρ∗, evaluated on the same validation set used for that selection. Because ρ∗ is chosen as the argmax of a score that includes fexec on these exact samples, the reported gain is entailed by the selection rule rather than being an out-of-sample prediction. The paper reports no test-set comparison of prompt levels; Table II evaluates only the optimized prompts on test sets, so the prompt-hierarchy improvement is not independently validated. This is a fitted input reported as a predictive gain.

full rationale

The central test-set comparison in Table II is empirically independent: the four VLFMs are evaluated on held-out test frames with fixed optimized prompts, and those accuracy numbers do not reduce to the APE objective. Self-citations to PIP-Net [4] and the context-aware review [24] exist but are not load-bearing, since they serve as baselines and related work rather than as justification for the prompt claims. However, the headline quantitative gains for vehicle-speed/time-conscious prompts and for APE optimization (19.8% and 12.5% in the abstract, 12.5% and 5.3% in the conclusion) are taken from Table I, which is produced on the same validation sets used to select the best prompts. Since the selected prompt is the argmax of a score containing validation accuracy, its reported accuracy gain over the baseline prompt is by construction the result of the optimization, not an independent prediction. No self-citation uniqueness chains, ansatz-smuggling via citations, or renaming of known results were found. The unresolved alpha parameter and the JAAD PDs/PDt annotation inconsistency are correctness risks rather than circularity. Overall, the prompt-improvement claims are partially circular, while the final VLM-versus-baseline ranking retains independent content.

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

No new physical entities or fitted constants are introduced. The main ledger entries are the prompt-scoring weight alpha and the APE hyperparameters, whose values are not reported. The evaluation relies on assumptions about validation-to-test transfer, baseline comparability, and API determinism.

free parameters (2)
  • alpha (weight in APE prompt scoring) = not reported
    Appears in Eq. (1) as the balance between execution accuracy and log probability. The chosen value changes which prompts APE selects, but no value or sensitivity analysis is given.
  • APE hyperparameters (M, T, K) = M=365, T=40, K=5
    Chosen by hand without reported sensitivity analysis; they govern the prompt optimization that produces the reported gains.
assumptions (3)
  • domain assumption Validation-set prompt selection transfers to test sets of the same three datasets without overfitting.
    Section III-B optimizes prompts on validation samples; Section III-D reports test results. With JAAD validation of 32 samples and FU-PIP validation of 90 samples, small accuracy gaps may be noise.
  • domain assumption The vision-based baseline numbers in Table II are directly comparable to the VLFMs' results.
    Section III-D compares against published baselines without re-running them under the same protocol, split, and metric definitions.
  • domain assumption Commercial VLFMs give stable, reproducible answers for a fixed prompt.
    Section III-D uses GPT-4V and GPT-4 mini via API; no temperature, seed, or repeated-run variability is reported, and API versions change over time.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pedestrian Intention Prediction via Vision-Language Foundation Models." pith.science (2026). https://pith.science/paper/Z5ZPKYUX

@misc{pith2026250704141,
  author       = {Pith},
  title        = {Pith review of: Pedestrian Intention Prediction via Vision-Language Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z5ZPKYUX}},
  note         = {Machine review of arXiv:2507.04141}
}
read the original abstract

Prediction of pedestrian crossing intention is a critical function in autonomous vehicles. Conventional vision-based methods of crossing intention prediction often struggle with generalizability, context understanding, and causal reasoning. This study explores the potential of vision-language foundation models (VLFMs) for predicting pedestrian crossing intentions by integrating multimodal data through hierarchical prompt templates. The methodology incorporates contextual information, including visual frames, physical cues observations, and ego-vehicle dynamics, into systematically refined prompts to guide VLFMs effectively in intention prediction. Experiments were conducted on three common datasets-JAAD, PIE, and FU-PIP. Results demonstrate that incorporating vehicle speed, its variations over time, and time-conscious prompts significantly enhances the prediction accuracy up to 19.8%. Additionally, optimised prompts generated via an automatic prompt engineering framework yielded 12.5% further accuracy gains. These findings highlight the superior performance of VLFMs compared to conventional vision-based models, offering enhanced generalisation and contextual understanding for autonomous driving applications.

Figures

Figures reproduced from arXiv: 2507.04141 by the authors.

Figure 1
Figure 1. Overview of vision-language crossing intention prediction model. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The measured performance of GPT-4V during the prompt [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 27 canonical work pages

  1. [1]

    Autonomous vehicles that interact with pedestrians: A survey of theory and practice,

    A. Rasouli and J. K. Tsotsos, “Autonomous vehicles that interact with pedestrians: A survey of theory and practice,” IEEE transactions on intelligent transportation systems , vol. 21, no. 3, pp. 900–918, 2019

    2019

  2. [2]

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

    H. Razali, T. Mordan, and A. Alahi, “Pedestrian intention prediction: A convolutional bottom-up multi-task approach,” Transportation re- search part C: emerging technologies , vol. 130, p. 103259, 2021

    work page 2021

  3. [3]

    St crossingpose: A spatial- temporal graph convolutional network for skeleton-based pedestrian crossing intention prediction,

    X. Zhang, P. Angeloudis, and Y . Demiris, “St crossingpose: A spatial- temporal graph convolutional network for skeleton-based pedestrian crossing intention prediction,” IEEE Transactions on Intelligent Trans- portation Systems, vol. 23, no. 11, pp. 20 773–20 782, 2022

    work page 2022

  4. [4]

    PIP-Net: Pedestrian Intention Prediction in the Wild

    M. Azarmi, M. Rezaei, H. Wang, and S. Glaser, “Pip-net: Pedestrian intention prediction in the wild,” arXiv preprint arXiv:2402.12810 , 2024

    work page Pith review arXiv 2024

  5. [5]

    Pedestrian action an- ticipation using contextual feature fusion in stacked rnns,

    A. Rasouli, I. Kotseruba, and J. K. Tsotsos, “Pedestrian action an- ticipation using contextual feature fusion in stacked rnns,” in British Machine Vision Conference, 2020

    work page 2020

  6. [6]

    Do they want to cross? understanding pedestrian intention for behavior prediction,

    I. Kotseruba, A. Rasouli, and J. K. Tsotsos, “Do they want to cross? understanding pedestrian intention for behavior prediction,” in 2020 IEEE Intelligent Vehicles Symposium (IV) , 2020, pp. 1688–1693

    work page 2020

  7. [7]

    Multi-modal hybrid architecture for pedestrian action prediction,

    A. Rasouli, T. Yau, M. Rohani, and J. Luo, “Multi-modal hybrid architecture for pedestrian action prediction,” in 2022 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2022, pp. 91–97

    work page 2022

  8. [8]

    Pedestrian graph+: A fast pedestrian crossing prediction model based on graph convo- lutional networks,

    P. R. G. Cadena, Y . Qian, C. Wang, and M. Yang, “Pedestrian graph+: A fast pedestrian crossing prediction model based on graph convo- lutional networks,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 11, pp. 21 050–21 061, 2022

    work page 2022

Show all 36 references
  1. [9]

    Visual reasoning using graph con- volutional networks for predicting pedestrian crossing intention,

    T. Chen, R. Tian, and Z. Ding, “Visual reasoning using graph con- volutional networks for predicting pedestrian crossing intention,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 3103–3109

    2021

  2. [10]

    CAPformer: Pedestrian crossing action prediction using transformer,

    J. Lorenzo, I. Parra, R. Izquierdo, A. L. Ballardini, ´A. Hern ´andez- Saz, D. F. Llorca, and M. ´A. Sotelo, “CAPformer: Pedestrian crossing action prediction using transformer,” Sensors (Basel, Switzerland) , vol. 21, 2021

    2021

  3. [11]

    Pit: Progressive interaction transformer for pedestrian crossing intention prediction,

    Y . Zhou, G. Tan, R. Zhong, Y . Li, and C. Gou, “Pit: Progressive interaction transformer for pedestrian crossing intention prediction,” IEEE Transactions on Intelligent Transportation Systems , 2023

    2023

  4. [12]

    Predicting pedestrian inten- tions with multimodal intentformer: A co-learning approach,

    N. Sharma, C. Dhiman, and S. Indu, “Predicting pedestrian inten- tions with multimodal intentformer: A co-learning approach,” Pattern Recognition, vol. 161, p. 111205, 2025

    2025

  5. [13]

    Pedestrian behavior inter- pretation from pose estimation,

    J. Yang, A. Gui, J. Wang, and J. Ma, “Pedestrian behavior inter- pretation from pose estimation,” in IEEE International Intelligent Transportation Systems Conference (ITSC) , 2021, pp. 3110–3115

    2021

  6. [14]

    Multi-scale pedestrian intent prediction using 3d joint information as spatio-temporal representation,

    S. Ahmed, A. Al Bazi, C. Saha, S. Rajbhandari, and M. N. Huda, “Multi-scale pedestrian intent prediction using 3d joint information as spatio-temporal representation,” Expert Systems With Applications , vol. 225, p. 120077, 2023

    2023

  7. [15]

    Spatiotemporal relationship reasoning for pedestrian intent prediction,

    B. Liu, E. Adeli, Z. Cao, K.-H. Lee, A. Shenoi, A. Gaidon, and J. C. Niebles, “Spatiotemporal relationship reasoning for pedestrian intent prediction,” IEEE Robotics and Automation Letters , vol. 5, no. 2, pp. 3485–3492, 2020

    2020

  8. [16]

    Real-time intent prediction of pedestrians for autonomous ground vehicles via spatio-temporal densenet,

    K. Saleh, M. Hossny, and S. Nahavandi, “Real-time intent prediction of pedestrians for autonomous ground vehicles via spatio-temporal densenet,” in 2019 International Conference on Robotics and Automa- tion (ICRA), 2019, pp. 9704–9710

    2019

  9. [17]

    Pedestrian-vehicle information modulation for pedestrian crossing intention prediction,

    L. Xu, S. You, G. He, and Y . Li, “Pedestrian-vehicle information modulation for pedestrian crossing intention prediction,” IEEE Trans- actions on Intelligent Vehicles , 2024

    2024

  10. [18]

    Causal reasoning in typical computer vision tasks,

    K. Zhang, Q. Sun, C. Zhao, and Y . Tang, “Causal reasoning in typical computer vision tasks,” Science China Technological Sciences, vol. 67, no. 1, pp. 105–120, 2024

    2024

  11. [19]

    Vision language models in autonomous driving: A survey and outlook,

    X. Zhou, M. Liu, E. Yurtsever, B. L. Zagar, W. Zimmer, H. Cao, and A. C. Knoll, “Vision language models in autonomous driving: A survey and outlook,” IEEE Transactions on Intelligent Vehicles, 2024

    2024

  12. [20]

    Gpt-4v takes the wheel: Promises and challenges for pedestrian behavior prediction,

    J. Huang, P. Jiang, A. Gautam, and S. Saripalli, “Gpt-4v takes the wheel: Promises and challenges for pedestrian behavior prediction,” in Proceedings of the AAAI Symposium Series , vol. 3, no. 1, 2024, pp. 134–142

    2024

  13. [21]

    Omnipredict: Gpt-4o enhanced multi-modal pedestrian crossing intention prediction

    J.-S. Ham, J. Huang, P. Jiang, J. Moon, Y . Kwon, S. Saripalli, and C. Kim, “Omnipredict: Gpt-4o enhanced multi-modal pedestrian crossing intention prediction.”

  14. [22]

    Pedvlm: Pedestrian vision language model for intentions prediction,

    F. Munir, S. Azam, T. Mihaylova, V . Kyrki, and T. P. Kucner, “Pedvlm: Pedestrian vision language model for intentions prediction,” 2024. [Online]. Available: https://openreview.net/forum?id=RAX45dcfA2

    2024

  15. [23]

    Cross or wait? predicting pedestrian interaction outcomes at unsignalized crossings,

    C. Zhang, A. H. Kalantari, Y . Yang, Z. Ni, G. Markkula, N. Merat, and C. Berger, “Cross or wait? predicting pedestrian interaction outcomes at unsignalized crossings,” in 2023 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2023, pp. 1–8

    2023

  16. [24]

    Feature importance in pedestrian intention prediction: A context-aware review,

    M. Azarmi, M. Rezaei, H. Wang, and A. Arabian, “Feature importance in pedestrian intention prediction: A context-aware review,” arXiv preprint arXiv:2409.07645, 2024

    2024 arXiv

  17. [25]

    Hierarchical prompting taxonomy: A universal evaluation framework for large language models,

    D. Budagam, S. KJ, A. Kumar, V . Jain, and A. Chadha, “Hierarchical prompting taxonomy: A universal evaluation framework for large language models,” arXiv preprint arXiv:2406.12644 , 2024

    2024 arXiv

  18. [26]

    Large language models are human-level prompt engineers,

    Y . Zhou, A. I. Muresanu, Z. Han, K. Paster, S. Pitis, H. Chan, and J. Ba, “Large language models are human-level prompt engineers,” arXiv preprint arXiv:2211.01910 , 2022

    2022 arXiv

  19. [27]

    Benchmark for evaluating pedestrian action prediction,

    I. Kotseruba and A. Rasouli, “Benchmark for evaluating pedestrian action prediction,” IEEE Winter Conference on Applications of Com- puter Vision (WACV), pp. 1257–1267, 2021

    2021

  20. [28]

    Better zero-shot reasoning with role-play prompting,

    A. Kong, S. Zhao, H. Chen, Q. Li, Y . Qin, R. Sun, X. Zhou, E. Wang, and X. Dong, “Better zero-shot reasoning with role-play prompting,” arXiv preprint arXiv:2308.07702 , 2023

    2023 arXiv

  21. [29]

    Good at captioning, bad at counting: Benchmarking gpt-4v on earth observation data,

    C. Zhang and S. Wang, “Good at captioning, bad at counting: Benchmarking gpt-4v on earth observation data,” arXiv preprint arXiv:2401.17600, 2024

    2024 arXiv

  22. [30]

    Is the pedestrian going to cross? answering by 2d pose estimation,

    Z. Fang and A. M. L ´opez, “Is the pedestrian going to cross? answering by 2d pose estimation,” in 2018 IEEE Intelligent Vehicles symposium (IV). IEEE, 2018, pp. 1271–1276

    2018

  23. [31]

    Chatgpt: Generative pre-trained transformer,

    OpenAI, “Chatgpt: Generative pre-trained transformer,” 2023, accessed: 2025-01-18. [Online]. Available: https://openai.com/chatgpt

    2023

  24. [32]

    Agreeing to cross: How drivers and pedestrians communicate,

    A. Rasouli, I. Kotseruba, and J. K. Tsotsos, “Agreeing to cross: How drivers and pedestrians communicate,” in IEEE Intelligent Vehicles Symposium (IV), 2017, pp. 264–269

    2017

  25. [33]

    PIE: A large-scale dataset and models for pedestrian intention estimation and trajectory prediction,

    A. Rasouli, I. Kotseruba, T. Kunic, and J. K. Tsotsos, “PIE: A large-scale dataset and models for pedestrian intention estimation and trajectory prediction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2019, pp. 6262–6271

    2019

  26. [34]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat et al. , “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774 , 2024

    2024 arXiv

  27. [35]

    Minigpt-4: En- hancing vision-language understanding with advanced large language models,

    D. Zhu, J. Chen, X. Shen, X. Li, and M. Elhoseiny, “Minigpt-4: En- hancing vision-language understanding with advanced large language models,” arXiv preprint arXiv:2304.10592 , 2023

    2023 arXiv

  28. [36]

    Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models,

    F. Li, R. Zhang, H. Zhang, Y . Zhang, B. Li, W. Li, Z. Ma, and C. Li, “Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models,” arXiv preprint arXiv:2407.07895 , 2024

    2024 arXiv

Pith tools

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