Pith. sign in

REVIEW 3 major objections 5 minor 25 references

TinyDrive: Multiscale Visual Question Answering with Selective Token Routing for Autonomous Driving

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TinyDrive, a 16M-31.8M parameter vision-language model, claims state-of-the-art BLEU-4 and METEOR scores on the DriveLM benchmark, beating models with up to 769M parameters.

desk verdict A compact driving VLM with a plausible efficiency story, but the SOTA claim is not supported by the under-specified evaluation and is contradicted by the paper's own Table 4. read the letter →

arxiv 2505.15564 v1 pith:CW2KH2GY submitted 2025-05-21 cs.CV

Hossein Hassani , Soodeh Nikan , Abdallah Shami This is my paper
classification cs.CV
keywords visualquestionansweringautonomousdrivingvisionlanguagemodelmultiscaleencodertokenroutingsequenceprioritybufferDriveLMbenchmarkefficientdeeplearning
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

TinyDrive is a compact vision-language model for visual question answering in autonomous driving, designed to run on resource-constrained vehicle hardware. The paper claims that, despite having only 16M or 31.8M parameters, TinyDrive outperforms larger published models on the DriveLM benchmark on BLEU-4 and METEOR, with relative gains of 11.1% and 35.4% over the previous best model. To get there the paper combines a multiscale CNN vision encoder with a dual-level prioritization scheme: a token routing mechanism that prunes less informative text tokens, and a sequence priority buffer that more frequently samples the most informative training sequences. A sympathetic reader would take the paper's central contribution to be evidence that aggressive parameter reduction does not have to cost language understanding accuracy in driving VQA.

What carries the argument

The two mechanisms that carry the argument are (1) a multiscale CNN vision encoder with scale injection, cross-scale gating, and a Local-Global Block that fuses high, mid, and low resolution features into a compact 144-dimensional embedding for six views, and (2) a dual-level prioritization: the token routing mechanism scores each text token by a weighted combination of embedding magnitude, a position weight that upweights middle tokens, and a padding mask, keeping only the top $K=64$ tokens; the sequence priority buffer scores each whole sequence by a weighted sum of normalized loss, uncertainty (inverse confidence), and diversity (cosine dissimilarity to other sequences), then samples from the buffer with importance-weight bias correction.

What would settle it

Run MiniDrive384, EM-VLM4ADBase, and TinyDrivev12 on the same DriveLM split with identical prompts, decoding parameters, and metric computation; if MiniDrive384's BLEU-4 or METEOR meets or exceeds TinyDrive's reported 55.82 or 50.56, the paper's state-of-the-art claim is false.

Watch

Extended reading notes

Core claim

The paper's central claim is that a deliberately tiny VLM can reach state-of-the-art language understanding on a driving VQA benchmark. On the DriveLM-nuScenes dataset, TinyDrivev12 (31.8M params) reports BLEU-4 55.82 and METEOR 50.56, a relative improvement of 11.1% and 35.4% over MiniDrive384, while using 76.8% fewer parameters; TinyDrivev11 (16M) also beats all compared baselines on BLEU-4 and METEOR. The authors attribute this to the combination of a multiscale CNN vision encoder and selective token/sequence prioritization rather than to larger language model capacity.

Load-bearing premise

The load-bearing premise is that the baseline scores cited from MiniDrive and EM-VLM4AD were obtained under the same evaluation protocol as TinyDrive's scores; the paper does not report the DriveLM split, prompt template, decoding settings, or metric computation used for either side.

Editorial extensions

If this is right

  • If the reported comparisons hold, TinyDrivev11 (16M parameters, 4.27B FLOPs) and TinyDrivev12 (31.8M parameters, 6.50B FLOPs) both outperform MiniDrive224/384 and EM-VLM4ADBase/Large on BLEU-4 and METEOR on DriveLM.
  • Averaged over BLEU-4, METEOR, ROUGE-L, and CIDEr, TinyDrivev12 achieves 44.73 versus 41.11 for MiniDrive384 and 38.75 for EM-VLM4ADBase.
  • The reported 99% FLOP reduction relative to EM-VLM4ADBase suggests that the dual-level prioritization scheme can make training and inference substantially cheaper without sacrificing the metrics that matter for driving QA.
  • Because T5-tiny and T5-mini are the language backbones, the approach demonstrates that a very small text model plus a carefully engineered vision encoder and selective token processing can support competitive driving VQA.

Reading between the lines

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

  • The claimed state-of-the-art scores depend on the baseline numbers being produced under the same evaluation protocol; the paper does not specify which DriveLM split, prompt template, decoding settings, or metric computation was used, so the headline comparison is the first thing to verify.
  • The token routing mechanism scores text tokens only by their own embedding properties, not by their relevance to the visual input; as the paper's Section 5 notes, extending the score to condition on visual content is a natural next step.
  • The sequence priority buffer, with its uncertainty and diversity terms, is a general training-prioritization idea that could transfer to other multimodal fine-tuning tasks beyond driving VQA.
  • If the efficiency claims replicate, the strongest implication is that parameter count is not a reliable proxy for VQA quality in driving domains; architecture and training-data selection may matter more.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents TinyDrive, a lightweight VLM for multi-view VQA in autonomous driving. It combines a multiscale CNN vision encoder with scale injection, cross-scale gating, and a local-global block, plus a dual-level prioritization scheme that routes top-K text tokens and reweights training sequences. The authors first evaluate two TinyDrive variants on a custom Rosmaster dataset and then compare against EM-VLM4AD and MiniDrive on DriveLM-nuScenes, claiming state-of-the-art BLEU-4 and METEOR with substantially fewer parameters.

Significance. If the reported results were reproducible under the official DriveLM protocol, TinyDrive would be a meaningful efficiency contribution: its parameter count (16M/31.8M) is an order of magnitude below the cited baselines, and the architecture is described in sufficient detail to be reimplemented. The authors are also candid in Section 5 about the limitation that token routing ignores cross-modal relevance. However, the paper's central SOTA claim is currently not assessable because the experimental comparison is under-specified and internally inconsistent. I do not see a circularity problem in using model-derived scores for prioritization; that is a standard active-learning loop.

major comments (3)
  1. [§4.2, Table 4] The manuscript does not specify which DriveLM-nuScenes split is used for evaluation, what prompt template is provided to TinyDrive and to the baselines, what decoding algorithm is used, or how the BLEU-4, METEOR, ROUGE-L, and CIDEr scores are computed. Since the headline SOTA claim rests entirely on Table 4, this omission is load-bearing. In addition, Table 4 itself contradicts the prose: TinyDrive v12 is worse than MiniDrive384 on ROUGE-L (69.50 vs. 73.50) and CIDEr (3.02 vs. 3.32), so the statement in §4.2 that TinyDrive shows 'superior performance on all metrics' is inaccurate. The authors should adopt the official DriveLM evaluation protocol on a named split, report all comparison conditions (prompts, decoding, metric computation), and include error bars; without this, the comparison cannot be verified.
  2. [§4.2, Table 5] The vision encoder used for DriveLM is listed with 'Fine-Tuning (vision)' set to N/A, yet this encoder was previously trained on the 165-image indoor Rosmaster dataset described in §4.1. The paper does not explain how the encoder is adapted to six-view outdoor nuScenes images, whether the Rosmaster-trained weights are used as initialization, or what visual features are fed to the T5 model. This gap makes the transfer results hard to interpret. Please clarify the initialization and any domain adaptation, and report an ablation using a randomly initialized or generic pretrained encoder.
  3. [§4.2, Figure 1] The 'average' metric used in Figure 1 and in the §4.2 discussion is computed by averaging BLEU-4, METEOR, ROUGE-L, and CIDEr, which have different scales and ranges (CIDEr is around 3 while the other metrics range roughly 40-70). This aggregation is therefore dominated by the first three metrics and is not a principled summary. The efficiency claim should be supported by a metric-by-metric comparison rather than an unweighted average of disparate scales.
minor comments (5)
  1. [Tables 2 and 4] The column header 'BLUE-4' should be 'BLEU-4'.
  2. [§3.3, Eq. (12)] Equation (12) uses lowercase 'l' in the condition 'L/4 ≤ l ≤ 3L/4' while the token index is denoted as ℓ elsewhere; please make the notation consistent.
  3. [Throughout] There are several typographical errors, including 'mutiscale' in the Introduction and 'Rosmater' in Table 5; a careful proofread is needed.
  4. [§4.1] The paper reports 100% accuracy on the Rosmaster vision-encoder validation but states only that 10% of the 2,344 images were used for validation; it would help to give the exact number of images in each split and to clarify the relationship between the 2,344 frames and the 165 selected frames.
  5. [Appendix B.1] The exact Hugging Face identifiers for 'T5-tiny' and 'T5-mini' should be given, since these are not standard T5 sizes and affect reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TinyDrive's architecture and training scores are design choices evaluated on an external benchmark, not predictions derived from their own definitions.

full rationale

No circular step is present in TinyDrive's derivation chain. The paper proposes an architecture and training scheme: a multiscale CNN vision encoder, a token routing score (Eq. 11), and a sequence priority score (Eq. 13), then evaluates the resulting model on the public DriveLM benchmark. These scoring functions are design choices, not quantities fitted to the benchmark and then renamed as predictions; the model's BLEU-4, METEOR, ROUGE-L, and CIDEr scores are obtained by training and inference, not by construction from the scoring definitions. The 100% Rosmaster accuracy is reported on the same custom dataset used to train the vision encoder's classification head, but it is not used as evidence for the central SOTA claim, which rests on the external DriveLM comparison. The self-citation [8] in the introduction is a routine reference to prior work and is not load-bearing. The main weakness of the paper is that the DriveLM evaluation protocol (split, prompt template, decoding, metric computation) is under-specified, making the SOTA comparison hard to reproduce; that is a correctness and reproducibility concern, not a circularity one. No equation in the paper reduces to its own input, and no fitted parameter is relabeled as a prediction.

Assumptions & free parameters 8 free parameters · 4 assumptions · 3 invented entities

The central claim depends on a set of heuristically chosen hyperparameters (alpha, beta, lambda weights, K) and several design assumptions about multiscale features, token informativeness, and sequence prioritization. None of the invented components is validated by ablations, and the free parameters are tuned on the same validation set used to report results.

free parameters (8)
  • alpha (prioritization exponent) = 0.6
    Controls sampling intensity; tuned by trial and error (Table 5).
  • beta (correction exponent) = 0.4
    Corrects prioritization bias; tuned and annealed (Table 5).
  • lambda_l (loss weight) = 0.5
    Sequence score weight; chosen on validation.
  • lambda_u (uncertainty weight) = 0.3
    Sequence score weight; chosen on validation.
  • lambda_d (diversity weight) = 0.2
    Sequence score weight; chosen on validation.
  • lambda_m (magnitude weight) = 1.0
    Token score weight; fixed a priori.
  • lambda_p (position weight) = 1.0
    Token score weight; fixed a priori.
  • K (top text tokens) = 64
    Number of retained tokens; set to 64 of 75 in all experiments.
assumptions (4)
  • domain assumption Pretrained T5 can be fine-tuned to answer driving questions accurately
    The language model is initialized from pretrained T5-tiny/mini and fine-tuned on VQA data (Section 3.1, B.3); no evidence is given that T5 is capable of this task beyond the reported results.
  • ad hoc to paper Multiscale CNN features capture both fine-grained and contextual visual information needed for VQA
    Design claim in Section 3.2; no prior theoretical or empirical demonstration is referenced.
  • ad hoc to paper Token score based on embedding magnitude, position, and padding mask reflects token informativeness
    The heuristic in Eq. (11) is proposed without empirical justification or ablation.
  • ad hoc to paper Sequence priority based on loss, uncertainty, and diversity identifies training examples that improve the model
    Used for sampling in Section 3.3, but never validated by removing the buffer or comparing with uniform sampling.
invented entities (3)
  • Local-Global Block (LGB)
    purpose: Combine local and global features for the classification head and vision embeddings
    Introduced in Section 3.2; its contribution is not isolated by ablation.
  • Scale Injection Mechanism (SIM)
    purpose: Generate multi-resolution features with channel/spatial attention
    Part of the vision encoder; no independent test.
  • Cross-Scale Gating Mechanism (CSGM)
    purpose: Adaptively fuse features from three scales
    Described in Section 3.2; effectiveness not separately evaluated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TinyDrive: Multiscale Visual Question Answering with Selective Token Routing for Autonomous Driving." pith.science (2026). https://pith.science/paper/CW2KH2GY

@misc{pith2026250515564,
  author       = {Pith},
  title        = {Pith review of: TinyDrive: Multiscale Visual Question Answering with Selective Token Routing for Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CW2KH2GY}},
  note         = {Machine review of arXiv:2505.15564}
}
read the original abstract

Vision Language Models (VLMs) employed for visual question-answering (VQA) in autonomous driving often require substantial computational resources that pose a challenge for their deployment in resource-constrained vehicles. To address this challenge, we introduce TinyDrive, a lightweight yet effective VLM for multi-view VQA in driving scenarios. Our model comprises two key components including a multiscale vision encoder and a dual-level prioritization mechanism for tokens and sequences. The multiscale encoder facilitates the processing of multi-view images at diverse resolutions through scale injection and cross-scale gating to generate enhanced visual representations. At the token level, we design a token routing mechanism that dynamically selects and process the most informative tokens based on learned importance scores. At the sequence level, we propose integrating normalized loss, uncertainty estimates, and a diversity metric to formulate sequence scores that rank and preserve samples within a sequence priority buffer. Samples with higher scores are more frequently selected for training. TinyDrive is first evaluated on our custom-curated VQA dataset, and it is subsequently tested on the public DriveLM benchmark, where it achieves state-of-the-art language understanding performance. Notably, it achieves relative improvements of 11.1% and 35.4% in BLEU-4 and METEOR scores, respectively, despite having a significantly smaller parameter count.

Figures

Figures reproduced from arXiv: 2505.15564 by the authors.

Figure 1
Figure 1. Comparison of TinyDrive against SOTA models on the DriveLM benchmark, evaluating [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of TinyDrive. modulates these visual tokens based on instructions extracted from text embeddings. The model achieves SOTA efficiency and superior performance on the DriveLM dataset. 3 Methodology In this section, we present the overall architecture of TinyDrive including detailed descriptions of the vision encoder, token routing mechanism, and priority buffer. This will then be followed by a tho… view at source ↗
Figure 3
Figure 3. Sample generated answers by TinyDrivev12 for VQA on the Rosmaster dataset. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The generated attention maps through the high, mid, and low resolution branches for two [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Sample images from the Rosmaster dataset. Images are captured using an RGB camera [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 5 canonical work pages

  1. [1]

    Meteor: An automatic metric for mt evaluation with improved correlation with human judgments

    Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. InProceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65–72, 2005

    2005

  2. [2]

    Driving with llms: Fusing object-level vector modality for explainable autonomous driving

    Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093–14100. IEEE, 2024

    2024

  3. [3]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019

    2019

  4. [4]

    Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition

    Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5513–5524, 2024

    work page 2024

  5. [5]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    arXiv 2010

  6. [6]

    Multi-frame, lightweight & effi- cient vision-language models for question answering in autonomous driving.arXiv preprint arXiv:2403.19838, 2024

    Akshay Gopalkrishnan, Ross Greer, and Mohan Trivedi. Multi-frame, lightweight & effi- cient vision-language models for question answering in autonomous driving.arXiv preprint arXiv:2403.19838, 2024

    arXiv 2024

  7. [7]

    World models for autonomous driving: An initial survey.IEEE Transac- tions on Intelligent Vehicles, 2024

    Yanchen Guan, Haicheng Liao, Zhenning Li, Jia Hu, Runze Yuan, Yunjian Li, Guohui Zhang, and Chengzhong Xu. World models for autonomous driving: An initial survey.IEEE Transac- tions on Intelligent Vehicles, 2024

    work page 2024

  8. [8]

    Traffic navigation via reinforcement learning with episodic-guided prioritized experience replay.Engineering Applications of Artificial Intelligence, 137:109147, 2024

    Hossein Hassani, Soodeh Nikan, and Abdallah Shami. Traffic navigation via reinforcement learning with episodic-guided prioritized experience replay.Engineering Applications of Artificial Intelligence, 137:109147, 2024

    work page 2024

Show all 25 references
  1. [9]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    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

    2022

  2. [10]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  3. [11]

    Rouge: A package for automatic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. InText summarization branches out, pages 74–81, 2004

    2004

  4. [12]

    Gpt-driver: Learning to drive with gpt.arXiv preprint arXiv:2310.01415, 2023

    Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt.arXiv preprint arXiv:2310.01415, 2023

    2023 arXiv

  5. [13]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002

    2002

  6. [14]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

    2021

  7. [15]

    Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020

    2020

  8. [16]

    Languagempc: Large language models as decision makers for autonomous driving

    H Sha, Y Mu, Y Jiang, L Chen, C Xu, P Luo, SE Li, M Tomizuka, W Zhan, and M Ding. Languagempc: Large language models as decision makers for autonomous driving. arxiv 2023. arXiv preprint arXiv:2310.03026, 2023. 10

    2023 arXiv

  9. [17]

    Drivelm: Driving with graph visual question answering

    Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beißwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. InEuropean Conference on Computer Vision, pages 256–274. Springer, 2024

    2024

  10. [18]

    Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  11. [19]

    Labelimg.https://github.com/tzutalin/labelImg, 2015

    Tzutalin. Labelimg.https://github.com/tzutalin/labelImg, 2015. Git code

    2015

  12. [20]

    Cider: Consensus-based image description evaluation

    Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4566–4575, 2015

    2015

  13. [21]

    Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving.arXiv preprint arXiv:2312.09245, 2023

    Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving.arXiv preprint arXiv:2312.09245, 2023

    2023

  14. [22]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  15. [23]

    Drivegpt4: Interpretable end-to-end autonomous driving via large language model.IEEE Robotics and Automation Letters, 2024

    Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model.IEEE Robotics and Automation Letters, 2024

    2024

  16. [24]

    Minidrive: More efficient vision-language models with multi-level 2d features as text tokens for autonomous driving

    Enming Zhang, Xingyuan Dai, Yisheng Lv, and Qinghai Miao. Minidrive: More efficient vision-language models with multi-level 2d features as text tokens for autonomous driving. arXiv preprint arXiv:2409.07267, 2024

    2024 arXiv

  17. [25]

    Vision language models in autonomous driving: A survey and outlook.IEEE Transactions on Intelligent Vehicles, 2024

    Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C Knoll. Vision language models in autonomous driving: A survey and outlook.IEEE Transactions on Intelligent Vehicles, 2024. A Rosmaster Dataset Rosmaster dataset was collected using ...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.