Pith. sign in

REVIEW 5 major objections 6 minor 34 references

STER-VLM: Spatio-Temporal With Enhanced Reference Vision-Language Models

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that a traffic-video VLM can beat direct fine-tuning without extra compute by decomposing captions into spatial-invariant and temporal-variant parts, selecting best-view frames, and adding reference hints.

desk verdict Honest engineering aggregation with a promising two-stage VQA gain; the captioning win is small, inconsistent, and the decomposition benefit is not isolated. read the letter →

arxiv 2508.13470 v1 pith:3NAZPWY3 submitted 2025-08-19 cs.CV cs.AI

Tinh-Anh Nguyen-Nhu , Triet Dao Hoang Minh , Dat To-Thanh , Phuc Le-Gia , Tuan Vo-Lan , Tien-Huy Nguyen This is my paper
classification cs.CVcs.AI
keywords vision-languagemodelstrafficvideocaptioningspatio-temporalunderstandingcaptiondecompositionLoRAfine-tuningvisualquestionansweringframeselectionsafetyanalysis
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 tries to show that a traffic-video vision-language model can be pushed past direct fine-tuning without extra compute by splitting its learning job. Ground-truth captions are decomposed into spatial-invariant and temporal-variant pieces; separate LoRA-tuned models specialize on each, and a composition model merges them. Frame inputs are chosen by temporal uniform sampling plus 'best-view' filtering based on bounding-box size, and a larger VLM supplies per-frame reference captions as hints. On the WTS and BDD validation sets the full pipeline raises the combined caption score from 29.975 to 30.258 and raises VQA accuracy from 75.367 to 81.507 through two-stage training. A score of 55.655 on the AI City Challenge 2025 Track 2 test set is reported as evidence the approach is resource-efficient.

What carries the argument

The load-bearing mechanism is the spatial/temporal caption decomposition pipeline: an LLM parses captions into spatial-invariant and temporal-variant components, two LoRA-tuned Qwen2.5-VL-7B specialists are trained on each component, and a composition model merges the outputs. Supporting mechanisms are best-view frame selection (largest bounding-box sums across cameras, with uniform sampling inside phases 4-5) and reference-driven prompts from a larger VLM. Together they convert fine-grained spatio-temporal differences into separable learning signals.

What would settle it

A direct check: take a held-out batch of WTS captions, run the same decomposition LLM on them, and have human judges measure whether the two parts, recombined, preserve the full semantic content of the original caption. If the CIDEr gain over baseline disappears when the decomposition is replaced by a random but equally sized sentence split, or if human evaluation finds no semantic preference for pipeline captions, the claimed gain in semantic richness would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that structured spatio-temporal decomposition of captions, rather than more parameters or more frames, is what lets a moderately sized VLM capture fine-grained traffic scenes. The pipeline parses each reference caption into a spatial-invariant part (appearance, environment) and a temporal-variant part (actions, positions, attention), fine-tunes two Qwen2.5-VL-7B specialists with LoRA, and then trains a composition model to merge their outputs. Frame selection draws three frames per phase, favoring cameras with the largest pedestrian/vehicle bounding boxes, and a larger VLM provides per-frame references inserted as hints. The paper reports that the combined caption score

Load-bearing premise

The load-bearing premise is that the LLM's decomposition of captions into spatial-invariant and temporal-variant parts is faithful enough to the WTS annotation structure that training on these splits teaches the model the right signal; if the decomposition errors are systematic, the semantic-richness gains could be an artifact of corrupted training targets.

Editorial extensions

If this is right

  • Two-stage LoRA training (caption first, then VQA) improves VQA accuracy from 75.367 to 81.507, implying caption generation is a useful intermediate task for visual question answering.
  • Caption decomposition trades strict n-gram overlap for semantic richness on the internal WTS subset, while on the external BDD subset all caption metrics improve.
  • Best-view filtering lets the model use fewer frames (three per phase) and still capture both appearance and action, reducing memory and compute relative to many-frame approaches.
  • Reference captions from a stronger VLM act as hints and improve the caption score from 30.061 to 30.257, without directly copying the reference text.
  • The full pipeline, trained on a single RTX A6000 GPU, reports 55.655 on the AI City Challenge 2025 Track 2 test leaderboard.

Reading between the lines

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

  • Because the headline gains are concentrated in CIDEr, which rewards rare but shared n-grams, the 'semantic richness' improvement could partly reflect the decomposition's tendency to repeat fragmented surface forms of the annotation style; a human evaluation or a paraphrase-invariant metric would tell whether true semantic coverage improves.
  • The reference-driven component could be strengthened by using confidence signals from the reference generator so that unreliable frames are not given misleading hints; the paper notes misinterpretations exist but does not quantify when they hurt.
  • The best-view criterion based on bounding-box area biases inputs toward large, close pedestrians; a direct test on small or occluded pedestrians would show whether the filtering mechanism generalizes across WTS's multi-view scenarios.
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 proposes STER-VLM, a framework for traffic-scene video captioning and visual question answering built on Qwen2.5-VL-7B with LoRA. The method combines (1) offline decomposition of captions into spatial-invariant and temporal-variant parts using Qwen2.5-72B-Instruct, (2) temporal frame sampling and best-view filtering, (3) frame-level reference captions from Qwen2.5-VL-72B used as hints, and (4) visual/textual prompt engineering. A composition model merges the two caption branches, and VQA is performed by continuing from captioning checkpoints. Experiments on WTS and BDD report caption metrics and VQA accuracy, and the framework scored 55.655 on AI City Challenge 2025 Track 2. The authors claim improved captioning and VQA while remaining computationally efficient.

Significance. The intended contribution—separating spatial and temporal caption factors while keeping training on a single GPU—is timely and potentially practical. The evaluation is not circular: the Qwen2.5-VL-72B references are used as auxiliary inputs rather than ground truth, and the main results use held-out validation/test splits, including an external leaderboard. These are strengths. However, the evidence as presented does not yet establish the central claims. The caption-decomposition component highlighted in the abstract is not isolated in Table 1, the aggregate gain is small, the WTS internal subset degrades, and no uncertainty quantification is provided. The caption-score formula is underspecified, and the efficiency claim is unmeasured. If the key components can be isolated and validated, the framework could be a useful efficient baseline, but the current manuscript is not sufficient.

major comments (5)
  1. [Sec. 4.3, Table 1] The headline claim that caption decomposition improves captioning is not supported by the presented comparison. The row 'P' differs from 'B' by the full pipeline (decomposition, frame selection, references, prompts), so Table 1 cannot isolate the decomposition. The combined caption score increases by only +0.283 (29.975→30.258), while the WTS subset, whose annotation structure the decomposition is designed to match, decreases from 29.160 to 28.595. The text acknowledges that decomposition lowers BLEU-4, METEOR, and ROUGE-L on WTS. No error bars or significance tests are reported. Please provide an ablation that toggles decomposition alone under otherwise identical inputs, with variance or paired tests, and report WTS/BDD separately.
  2. [Sec. 3.1] The decomposition stage is a correctness-risk point. Qwen2.5-72B-Instruct produces the training targets for the spatial and temporal models, but the manuscript gives no validation of fidelity: no human agreement, no comparison against WTS annotation structure, no error analysis. If the LLM paraphrases or omits information, the fine-tuned models learn from corrupted targets; the CIDEr improvement could then reflect LLM-style phrasing rather than improved spatio-temporal understanding. Given Sec. 4.3 reports a drop in n-gram metrics on WTS, please validate the decomposition on a sample and report per-component metric changes, not only the aggregated caption score.
  3. [Sec. 4.2.1] The caption score is used throughout Tables 1–5 but is not defined exactly. 'Full weight to BLEU-4, METEOR, and ROUGE-L, and a lower weight to CIDEr' is insufficient. Without the formula and weights, the reported deltas (e.g., +0.283, −0.565) cannot be interpreted, and the results are not reproducible. Please specify the exact scoring formula and, ideally, release the evaluation script.
  4. [Sec. 3.2 / Sec. 4.3] Temporal frame selection and best-view filtering are presented as a contribution, but no quantitative ablation supports them. Figure 3 shows an anecdotal example of first-frame-only failure, and the text says prior works use only the first frame, but no table reports captioning scores with first-frame-only vs. the proposed selection. Without this comparison, the effect of frame selection is unquantified. Add an ablation using the same model and inputs except for the frame-selection strategy.
  5. [Abstract / Sec. 3.5 / Table 6] The efficiency claim is not evidenced. The paper says the framework is computationally efficient, but no runtime, memory, FLOPs, or comparison with the multi-model Divide-and-Conquer baseline [26] is given. Similarly, Table 6 reports a VQA gain from two-stage training, but the direct-finetuning setup is not described and no variance or significance is reported. Please provide resource measurements and more experimental detail for the VQA comparison.
minor comments (6)
  1. [Sec. 3.1] The sentence beginning 'we input captions from all temporal, and since spatial-invariant characteristics remain consistent across phases...' is garbled/incomplete and does not parse. Please rewrite.
  2. [Sec. 3.2, Eq. (1)] The selection rule and its legend are typeset with missing symbols (placeholders such as '�phase', '�����', '���������'); the formula is not readable. Define all variables and fix the typesetting.
  3. [Sec. 3.4.1] 'based on the given coordination' should be 'based on the given coordinates'.
  4. [Sec. 3.5] Typo: 'comprehensice' should be 'comprehensive'.
  5. [Fig. 3 caption] The caption says the first two frames capture the pedestrian's appearance while the last two provide environmental context, which appears inconsistent with the surrounding text; please check and align.
  6. [Sec. 2.1] Typographical and formatting issues: 'desgined', 'novels dataset', and inconsistent citation formatting (e.g., [4, 23, 26, 32]) should be cleaned up.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the system's claims are evaluated against held-out WTS/BDD annotations and the official AI City Challenge leaderboard, and no prediction reduces to a fitted input or self-citation chain.

full rationale

The paper is an empirical systems paper with no mathematical derivation chain that could collapse into its inputs. The caption-decomposition step (Sec. 3.1) uses Qwen2.5-72B-Instruct as an external parser to split ground-truth captions into spatial-invariant and temporal-variant parts; these splits are used as training targets, while final captions are evaluated against the original human references. Reference-driven hints (Sec. 3.3) are generated by Qwen2.5-VL-72B and injected as prompt inputs, not as ground-truth labels; Table 2 measures the difference with and without these hints on validation data. The captioning and VQA results in Tables 1 and 6 are computed on held-out validation/test sets, and Table 7 compares against an external leaderboard. The self-citations [13,14,15,21] appear only in the introduction as background and are not load-bearing for any experimental claim. The paper explicitly concedes in Sec. 4.3 that decomposition disrupts exact phrase structure on the internal WTS subset, lowering BLEU-4, METEOR, and ROUGE-L, and that the fixed decomposition limits merge flexibility; this is an acknowledged trade-off and a validity/correctness concern, but it does not make any result equivalent to its inputs by construction. Similarly, Table 1 compares the full pipeline against an enhanced-prompt baseline, so the decomposition benefit is not isolated; this is an experimental-design weakness, not circularity. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from prior work, and the teacher-student same-family concern is not a circularity because the teacher outputs are auxiliary inputs, not the evaluation target.

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

The paper introduces no new physical or formal entities. The load-bearing choices are empirical design decisions: frame count, view-selection criteria, manually curated prompt hints, and reliance on LLM decomposition and teacher references. These are not derived from theory and are only partially ablated.

free parameters (4)
  • frames per phase = 3
    Hand-chosen in Section 3.2 to balance richness and compute; no quantitative sweep is reported.
  • best-view selection criteria = largest pedestrian bbox; largest sum of pedestrian+vehicle bboxes
    Hand-designed rules in Section 3.2; no ablation table isolates their effect.
  • LoRA rank = 64
    Chosen without a reported sweep in Section 4.2.2; affects capacity and results.
  • textual hint groups = Appearance, Environment, Action, Attention, Location
    Manually curated in Section 3.4.2 and ablated only by removing whole groups.
assumptions (4)
  • domain assumption Qwen2.5-72B-Instruct decomposition of captions into spatial-invariant and temporal-variant parts is faithful and matches the WTS annotation structure.
    Section 3.1 uses the LLM output as training targets but provides no validation of parse accuracy.
  • domain assumption Three uniformly sampled frames per phase plus best-view filtering captures the temporal action progression of each phase.
    Section 3.2 assumes phase boundaries mark behavioral transitions and that first, middle, and last frames summarize the action; no quantitative comparison to other sampling rules is given.
  • domain assumption Reference captions generated by Qwen2.5-VL-72B are useful hints and errors are ignored by the student model.
    Section 3.3 and Figure 5 admit the teacher sometimes misinterprets details; the model is assumed to filter these errors.
  • domain assumption Validation scores on WTS and BDD transfer to the official AI City Challenge 2025 Track 2 test set.
    Used in Section 4.3 and Table 7; no analysis of distribution shift is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STER-VLM: Spatio-Temporal With Enhanced Reference Vision-Language Models." pith.science (2026). https://pith.science/paper/3NAZPWY3

@misc{pith2026250813470,
  author       = {Pith},
  title        = {Pith review of: STER-VLM: Spatio-Temporal With Enhanced Reference Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3NAZPWY3}},
  note         = {Machine review of arXiv:2508.13470}
}
read the original abstract

Vision-language models (VLMs) have emerged as powerful tools for enabling automated traffic analysis; however, current approaches often demand substantial computational resources and struggle with fine-grained spatio-temporal understanding. This paper introduces STER-VLM, a computationally efficient framework that enhances VLM performance through (1) caption decomposition to tackle spatial and temporal information separately, (2) temporal frame selection with best-view filtering for sufficient temporal information, and (3) reference-driven understanding for capturing fine-grained motion and dynamic context and (4) curated visual/textual prompt techniques. Experimental results on the WTS \cite{kong2024wts} and BDD \cite{BDD} datasets demonstrate substantial gains in semantic richness and traffic scene interpretation. Our framework is validated through a decent test score of 55.655 in the AI City Challenge 2025 Track 2, showing its effectiveness in advancing resource-efficient and accurate traffic analysis for real-world applications.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 27 canonical work pages

  1. [26]

    SUTD-TrafficQA: A Ques- tion Answering Benchmark and an Efficient Network for Video Reasoning Over Traffic Events

    Li Xu, He Huang, and Jun Liu. SUTD-TrafficQA: A Ques- tion Answering Benchmark and an Efficient Network for Video Reasoning Over Traffic Events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9878–9888, 2021. 2

    work page 2021

  2. [1]

    Qwen-vl: A frontier large vision-language model with versatile abilities

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1

    arXiv 2023

  3. [2]

    Qwen2.5-vl technical report, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao- hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report,...

    work page 2025

  4. [3]

    Maplm: A real-world large-scale vision-language benchmark for map and traffic scene un- derstanding

    Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. Maplm: A real-world large-scale vision-language benchmark for map and traffic scene un- derstanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21819– 21830, 2024. 2

    work page 2024

  5. [4]

    Cityllava: Efficient fine- tuning for vlms in city scenario

    Zhizhao Duan, Hao Cheng, Duo Xu, Xi Wu, Xiangxie Zhang, Xi Ye, and Zhen Xie. Cityllava: Efficient fine- tuning for vlms in city scenario. In 2024 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition Work- shops (CVPRW), pages 7180–7189, 2024. 1, 4, 6

    work page 2024

  6. [5]

    Cityllava: Efficient fine-tuning for vlms in city scenario

    Zhizhao Duan, Hao Cheng, Duo Xu, Xi Wu, Xiangxie Zhang, Xi Ye, and Zhen Xie. Cityllava: Efficient fine-tuning for vlms in city scenario. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7180–7189, 2024. 2

    work page 2024

  7. [6]

    Optimal gradient checkpoint search for arbitrary computation graphs

    Jianwei Feng and Dong Huang. Optimal gradient checkpoint search for arbitrary computation graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11433–11442, 2021. 5

    work page 2021

  8. [7]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3, 4, 5

    work page 2022

Show all 34 references
  1. [8]

    Better zero-shot reasoning with role-play prompting, 2024

    Aobo Kong, Shiwan Zhao, Hao Chen, Qicheng Li, Yong Qin, Ruiqi Sun, Xin Zhou, Enzhi Wang, and Xiaohang Dong. Better zero-shot reasoning with role-play prompting, 2024. 4

    2024

  2. [9]

    Wts: A pedestrian-centric traffic video dataset for fine-grained spatial-temporal understand- ing

    Quan Kong, Yuki Kawana, Rajat Saini, Ashutosh Kumar, Jingjing Pan, Ta Gu, Yohei Ozao, Balazs Opra, Yoichi Sato, and Norimasa Kobori. Wts: A pedestrian-centric traffic video dataset for fine-grained spatial-temporal understand- ing. In European Conference on Computer Vision, pages 1–

  3. [10]

    Llava-st: A multimodal large language model for fine-grained spatial- temporal understanding

    Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaoming Wei, and Si Liu. Llava-st: A multimodal large language model for fine-grained spatial- temporal understanding. arXiv preprint arXiv:2501.08282 ,

    arXiv

  4. [11]

    Surveillancevqa-589k: A benchmark for comprehensive surveillance video-language understanding with large mod- els

    Bo Liu, Pengfei Qiao, Minhan Ma, Xuange Zhang, Yi- nan Tang, Peng Xu, Kun Liu, and Tongtong Yuan. Surveillancevqa-589k: A benchmark for comprehensive surveillance video-language understanding with large mod- els. arXiv preprint arXiv:2505.12589, 2025. 2

    2025 arXiv

  5. [12]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26296–26306, 2024. 1, 3

    2024

  6. [13]

    Improving generalization in visual reasoning via self-ensemble, 2024

    Tien-Huy Nguyen, Quang-Khai Tran, and Anh-Tuan Quang- Hoang. Improving generalization in visual reasoning via self-ensemble, 2024. 1

    2024

  7. [14]

    Hybrid, unified and itera- tive: A novel framework for text-based person anomaly re- trieval

    Tien-Huy Nguyen, Huu-Loc Tran, Huu-Phong Phan- Nguyen, and Quang-Vinh Dinh. Hybrid, unified and itera- tive: A novel framework for text-based person anomaly re- trieval. In Companion Proceedings of the ACM on Web Con- ference 2025, page 1576–1580, New York, NY , USA, 2025. Ass...

    2025

  8. [15]

    Le, and Quang-Vinh Dinh

    Tinh-Anh Nguyen-Nhu, Huu-Loc Tran, Nguyen-Khang Le, Minh-Nhat Nguyen, Tien-Huy Nguyen, Hoang-Long Nguyen-Huu, Huu-Phong Phan-Nguyen, Huy-Thach Pham, Quan Nguyen, Hoang M. Le, and Quang-Vinh Dinh. A lightweight moment retrieval system with global re-ranking and robust adaptive ...

    2025

  9. [16]

    Gpt-4v(ision)

    OpenAI. Gpt-4v(ision). 2023. 1

    2023

  10. [17]

    Roadsocial: A diverse videoqa dataset and benchmark for road event understanding from so- cial video narratives

    Chirag Parikh, Deepti Rawat, Tathagata Ghosh, Ravi Ki- ran Sarvadevabhatla, et al. Roadsocial: A diverse videoqa dataset and benchmark for road event understanding from so- cial video narratives. In Proceedings of the Computer Vision and Pattern Recognition Conference , pages ...

  11. [18]

    1, 2, 3, 5

    Springer, 2024. 1, 2, 3, 5

    2024

  12. [19]

    Safeplug: Empow- ering multimodal llms with pixel-level insight and temporal grounding for traffic accident understanding, 2025

    Zihao Sheng, Zilin Huang, Yen-Jung Chen, Yansong Qu, Yuhao Luo, Yue Leng, and Sikai Chen. Safeplug: Empow- ering multimodal llms with pixel-level insight and temporal grounding for traffic accident understanding, 2025. 2

    2025

  13. [20]

    Scvlm: Enhancing vision-language model for safety-critical event understanding, 2025

    Liang Shi, Boyu Jiang, Tong Zeng, and Feng Guo. Scvlm: Enhancing vision-language model for safety-critical event understanding, 2025. 2

    2025

  14. [21]

    What does clip know about a red circle? vi- sual prompt engineering for vlms

    Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? vi- sual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 11987–11997, 2023. 4

    2023

  15. [22]

    Le, and Quang- Vinh Dinh

    Huu-Loc Tran, Tinh-Anh Nguyen-Nhu, Huu-Phong Phan- Nguyen, Tien-Huy Nguyen, Nhat-Minh Nguyen-Dich, Anh Dao, Huy-Duc Do, Quan Nguyen, Hoang M. Le, and Quang- Vinh Dinh. Towards efficient and robust moment retrieval system: A unified framework for multi-granularity models and te...

    2025

  16. [23]

    Accidentgpt: A v2x environmental perception multi-modal large model for acci- dent analysis and prevention

    Lening Wang, Yilong Ren, Han Jiang, Pinlong Cai, Daocheng Fu, Tianqi Wang, Zhiyong Cui, Haiyang Yu, Xuesong Wang, Hanchu Zhou, et al. Accidentgpt: A v2x environmental perception multi-modal large model for acci- dent analysis and prevention. In 2024 IEEE Intelligent Vehi- cles...

    2024

  17. [24]

    Anastasiu, Zheng Tang, Ming- Ching Chang, Yue Yao, Liang Zheng, Mohammed Shaiqur Rahman, Meenakshi S

    Shuo Wang, David C. Anastasiu, Zheng Tang, Ming- Ching Chang, Yue Yao, Liang Zheng, Mohammed Shaiqur Rahman, Meenakshi S. Arya, Anuj Sharma, Pranamesh Chakraborty, Sanjita Prajapati, Quan Kong, Norimasa Ko- bori, Munkhjargal Gochoo, Munkh-Erdene Otgonbold, Gan- zorig Batnasan,...

    2024

  18. [25]

    Cogvlm: Visual expert for pretrained language models

    Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Keqin Chen, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. In Advances in Neural Infor...

    2024

  19. [27]

    Di- vide and conquer boosting for enhanced traffic safety de- scription and analysis with large vision language model

    Khai Trinh Xuan, Khoi Nguyen Nguyen, Bach Hoang Ngo, Vu Dinh Xuan, Minh-Hung An, and Quang-Vinh Dinh. Di- vide and conquer boosting for enhanced traffic safety de- scription and analysis with large vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vi...

    2024

  20. [28]

    Cpt: Colorful prompt tuning for pre-trained vision-language models

    Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, Tat seng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. ArXiv, abs/2109.11797, 2021. 4

    2021 arXiv

  21. [29]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar- rell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) ,

  22. [30]

    A study of sit- uational reasoning for traffic understanding

    Jiarui Zhang, Filip Ilievski, Kaixin Ma, Aravinda Kollaa, Jonathan Francis, and Alessandro Oltramari. A study of sit- uational reasoning for traffic understanding. arXiv preprint arXiv:2306.02520, 2023. 2

    2023 arXiv

  23. [31]

    When language and vi- sion meet road safety: leveraging multimodal large lan- guage models for video-based traffic accident analysis.arXiv preprint arXiv:2501.10604, 2025

    Ruixuan Zhang, Beichen Wang, Juexiao Zhang, Zilin Bian, Chen Feng, and Kaan Ozbay. When language and vi- sion meet road safety: leveraging multimodal large lan- guage models for video-based traffic accident analysis.arXiv preprint arXiv:2501.10604, 2025. 2

    2025 arXiv

  24. [32]

    Crashsage: A large language model-centered framework for contextual and interpretable traffic crash analysis

    Hao Zhen and Jidong J Yang. Crashsage: A large language model-centered framework for contextual and interpretable traffic crash analysis. arXiv preprint arXiv:2505.07853 ,

    arXiv

  25. [33]

    Tumtraffic-videoqa: A benchmark for unified spatio- temporal video understanding in traffic scenes

    Xingcheng Zhou, Konstantinos Larintzakis, Hao Guo, Wal- ter Zimmer, Mingyu Liu, Hu Cao, Jiajie Zhang, Venkat- narayanan Lakshminarasimhan, Leah Strand, and Alois C Knoll. Tumtraffic-videoqa: A benchmark for unified spatio- temporal video understanding in traffic scenes. arXiv ...

    2025 arXiv

  26. [34]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models, 2025

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shen- glong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, Hao Li, Jiahao Wang, Nianchen Deng, Songze L...

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.