Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

RoadMamba: A Dual Branch Visual State Space Model for Road Surface Classification

T0 review · 4 major / 5 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 RoadMamba claims 92.81% top-1 accuracy on a one-million-image road surface benchmark, beating MambaVision and LocalMamba.

desk verdict A competent dual-branch SSM architecture for road surface classification whose SOTA claim is undercut by an omitted prior baseline and an under-specified training protocol. read the letter →

arxiv 2508.01210 v1 pith:547YH2LF submitted 2025-08-02 cs.CV

Tianze Wang , Zhang Zhang , Chao Yue , Nuoran Li , Chao Sun This is my paper · ORCID
classification cs.CV
keywords roadsurfaceclassificationvisualstatespacemodelMambadual-brancharchitectureglobal-localfeaturefusionattentionautonomousdrivingfine-grained
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 a state-space-model (Mamba) backbone can become the best published classifier for fine-grained road surface conditions, provided it combines a global scan with a dedicated local-texture stream. It introduces RoadMamba, built from DualSSM blocks that scan the full image and randomly sampled 7×7 windows in parallel, fuse the two streams with channel and spatial attention, and add branch-specific auxiliary losses during training. On the RSCD dataset of one million images labelled into 27 road-condition classes, RoadMamba-B reports 92.81% top-1 accuracy, ahead of MambaVision-B at 92.07% and LocalMamba-S at 92.05%. A companion experiment reports that a windowed Mamba variant beats Swin Transformer by 3.8 points, which the authors take as evidence that pure Mamba scanning is preferable to window attention for this task.

What carries the argument

The central object is the DualSSM block. It contains GlobalSSM, which flattens the feature map row-major and column-major into two long one-dimensional sequences and runs shared selective state-space units, and LocalSSM, which partitions the map into 7×7 windows and scans only a randomly selected half of them, both horizontally and vertically. Dual Attention Fusion then applies channel attention to the global stream and spatial attention to the local stream, and sums the recalibrated maps. Auxiliary heads on each branch add cross-entropy losses with weight λ=0.3 during training only. The block's job is to keep global semantics and local texture available at every stage and to prevent the deep network from ignoring local cues.

What would settle it

Retrain Swin-B, ConvNeXt-B, ViT-B, MambaVision-B, and LocalMamba-S using RoadMamba's exact training recipe (same 40 epochs, batch size 32, AdamW, cosine schedule, weight decay, augmentation, and initialization) on the same split of RSCD; if any baseline reaches or exceeds 92.81% top-1 accuracy, the paper's state-of-the-art claim as stated would not survive the comparison.

Watch

Extended reading notes

Core claim

The central discovery is that the obstacle to state-of-the-art road surface classification with Mamba is not the global receptive field but the loss of local texture, and that a parallel local branch plus attention fusion fixes it. The paper's strongest evidence is Table 1, where RoadMamba-B reaches 92.81% top-1 accuracy versus ConvNeXt-B 84.08%, Swin-B 85.68%, ViT-B 86.83%, VMamba-B 91.11%, MambaVision-B 92.07%, and LocalMamba-S 92.05% on the same RSCD benchmark. The paper also reports that removing the dual scan, the Dual Attention Fusion, or the auxiliary loss drops accuracy (90.46 to 91.52 in the 20-epoch ablation), and that a windowed Mamba without the global branch achieves 89.22% versus Swin-T's 85.39%.

Load-bearing premise

The load-bearing assumption is that every model in the comparison table was trained and evaluated under the same, equally well-tuned protocol on the same RSCD split, even though the paper does not report learning rate, warmup, data augmentation, or initialization for any model.

Editorial extensions

If this is right

  • If the reported results hold, a pure state-space backbone with a parallel local branch is enough to top a 1-million-sample road surface benchmark, making Mamba architectures a viable choice for on-board perception systems.
  • The dual auxiliary loss is a training-time mechanism for preserving local texture; it adds no inference cost because the auxiliary heads are removed at test time.
  • The WindowMamba-versus-Swin result (89.22% vs 85.39% for Tiny) implies that, on this dataset, replacing window attention with windowed state-space scans is beneficial even without a global branch.
  • RoadMamba sets new reference numbers on RSCD across top-1 accuracy, mean precision, mean recall, and mean F1, which downstream planning controllers could rely on for better road-condition input under the paper's training protocol.

Reading between the lines

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

  • The reported 0.74-point edge over MambaVision-B is only meaningful if every baseline was tuned equally; the paper does not state learning rate, warmup, augmentation, or whether ImageNet weights were used, so the true margin under matched tuning is untested.
  • Because the local branch randomly drops half its windows, it doubles as a regularizer; one testable extension is measuring whether RoadMamba degrades more gracefully than Swin or VMamba on images with partial occlusion or local noise.
  • The WindowMamba-versus-Swin gap (89.22% vs 85.39%) is larger than typical architecture gaps and suggests protocol sensitivity; a replication with equal augmentation and learning-rate schedules would confirm or weaken the claim that pure Mamba beats window attention.
  • The same dual global/local design could transfer to other fine-grained visual tasks, such as pavement defect detection or tire-road friction estimation, where local texture carries the label.
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

4 major / 5 minor

Summary. The paper proposes RoadMamba, a dual-branch visual state space model for road surface classification on the RSCD dataset, which contains one million images with 27 combined classes of friction, material, and unevenness. The architecture combines a global SSM branch and a local windowed SSM branch, fuses them with a Dual Attention Fusion (DAF) module, and adds auxiliary classification losses on each branch to prevent the model from relying solely on the global branch. The authors report top-1 accuracies of 92.58/92.63/92.81 for the T/S/B variants on RSCD, claiming state-of-the-art performance over ConvNeXt, Swin, ViT, VMamba, LocalMamba, and MambaVision, and present ablations of the dual-scan strategy, DAF, and auxiliary loss. The paper also includes a motivating comparison between a window-based Mamba variant and Swin Transformer.

Significance. If the empirical results are correct and reproducible, the paper demonstrates a clean and effective way to combine global and local receptive fields in a Mamba backbone, with a plausible mechanism (branch-wise auxiliary losses) for preserving local texture information. The architectural design is novel in the road-surface-classification context, and the ablation order is internally consistent with the prose. The central claims are falsifiable benchmark numbers on a large public dataset, which is a strength. However, the significance of the SOTA claim depends on the completeness and fairness of the comparison, which the current manuscript does not fully establish.

major comments (4)
  1. [Related Work; Table 1] The paper's own prior method RoadFormer (Wang, Zhang, and Sun 2025) is cited in the Related Work as an existing RSCD approach, yet its accuracy is omitted from Table 1. Since the abstract and conclusion claim state-of-the-art performance, the number to beat is the previous best published RSCD result. If RoadFormer reports an accuracy at or above the 92.81% of RoadMamba-B, the headline claim is unsupported as stated. Please add RoadFormer to Table 1, or explicitly justify its exclusion with a concrete reason (e.g., different train/test split), and adjust the SOTA claim accordingly.
  2. [Implementation Details; Table 2] The training protocol is under-specified: the paper gives only the optimizer (AdamW), weight decay, epsilon, momentum, batch size, and a vague 'cosine annealing with an initial linear warmup phase.' It does not state the base learning rate, warmup length, data augmentation, whether ImageNet-pretrained weights are used, or how many seeds are averaged. The 3.8-point gap between WindowMamba-T (89.22) and Swin-T (85.39) in Table 2 is much larger than expected for two local-window models of similar size and suggests substantial differences in training recipe or initialization. Because RoadMamba-B beats MambaVision-B by only 0.74%, a modest baseline-tuning difference could erase the reported superiority in Table 1. Please report the full protocol and, ideally, rerun the baselines under an identical, carefully tuned recipe.
  3. [DualSSM Block, Eq. (9)] The local branch randomly selects only half of the windows for SSM scanning and sets the others to zero. The text says this acts 'similar to Dropout' during training, but it never states what happens at inference. If the same random selection is used at inference, the model is stochastic and the reported accuracies are not deterministic; if random selection is disabled, the local branch sees a different input distribution at inference, which should be described and empirically validated. This is load-bearing because the local branch contributes zero features for half the spatial locations, and the DAF spatial attention cannot restore them. Please specify the inference-time behavior and, if different from training, provide an explicit sentence and ideally an ablation.
  4. [Tables 1 and 3] All experimental results are single runs with no error bars or significance tests. The ablation improvements in Table 3 are small (0.16%, 0.73%, 0.17%), and the main 40-epoch comparison in Table 1 rests on a single trial per model. For a benchmark SOTA claim, margins of 0.16% or 0.74% are within the typical run-to-run variance of deep networks on large classification tasks. Please report at least three seeds with mean and standard deviation, or conduct a significance test, to support the claims that the dual-scan strategy, DAF, and auxiliary loss each help, and that RoadMamba-B is superior to MambaVision-B and LocalMamba-S.
minor comments (5)
  1. [Experiments, first paragraph] There is a typo: 'InTable1' should read 'In Table 1.'
  2. [Table 3 vs. Table 1] Table 3 is trained for 20 epochs while Table 1 is trained for 40 epochs; the text should state this explicitly in the ablation section, as the same model variant (RoadMamba) shows 91.52% in Table 3 and 92.58% (RoadMamba-T) in Table 1, which may otherwise confuse readers.
  3. [DualSSM Block; Figure 1; Appendix Table 6] The main text refers to '1-D SSM units' and 'SSM_h/SSM_v', while the appendix uses 'SS2D(Global Scan)' and 'SS2D(Local Scan)'. Please unify the terminology across the paper.
  4. [Figure 2 caption] The caption for panel (d) says 'performs bidirectional scanning separately in global and local windows,' but the figure shows only a single label for the RoadMamba scan pattern; please clarify which arrows correspond to the global scan and which to the local scan.
  5. [Why Choose Mamba?; Table 2] The WindowMamba vs. Swin comparison is not a clean ablation: the two models differ in many details beyond the attention-versus-SSM choice (e.g., window shift, positional encoding, normalization, and the random window-zeroing strategy). The concluding sentence that Mamba is 'a robust alternative to attention-based mechanisms' is too strong given this single uncontrolled comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: RoadMamba's accuracy is an externally measured test-set result; RoadFormer omission and protocol gaps are fairness issues, not definitional reductions.

full rationale

The paper's central claim—RoadMamba reaches 92.81% top-1 accuracy on RSCD—is a measured test-set number, not a quantity derived from fitted parameters or from a self-citation. The architecture components (DualSSM global/local scans, DAF attention fusion, auxiliary losses) are evaluated by ablations in Tables 3–5 and by direct comparison with independently published baselines; each component's contribution is an incremental accuracy difference on held-out labels. No equation in the paper defines an output in terms of the target accuracy, and no fitted hyperparameter (e.g., λ=0.3) is renamed as a prediction. The self-citations (RoadFormer, Pillarmamba, HeightFormer) are background or prior-work references; RoadFormer's absence from Table 1 is a benchmark-completeness and fairness concern that weakens the empirical force of the SOTA claim, but it is not circular—the paper never uses the RoadFormer citation to define or derive RoadMamba's accuracy. Likewise, the under-specified training protocol (no base learning rate, warmup length, augmentation, pretraining, or seeds) is a reproducibility and fairness risk, not a definitional reduction. No uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via citation. The derivation chain is therefore self-contained with respect to circularity, even though the empirical support for the SOTA claim is incomplete.

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

The central claim rests on the RSCD benchmark and its labels, standard SSM discretization math, and a small set of hand-set hyperparameters (lambda=0.3, window size 7, reduction ratio 4, 50% window sampling). No new physical or conceptual entities are postulated; DualSSM and DAF are network modules, not entities in the ledger sense. The unquantified load-bearing burden is the fairness of the comparison protocol: the paper omits the learning rate, pretraining status, and data split, so all SOTA conclusions inherit that assumption.

free parameters (4)
  • Auxiliary loss weight lambda = 0.3
    Balance coefficient in the total loss (Eq. 11), set by hand; controls how strongly the global and local branch auxiliary losses contribute.
  • Local window size M = 7
    Window size for the local scanning branch, 'set to 7 by default' in the DualSSM section; inherited from Swin/LocalMamba conventions rather than tuned on RSCD.
  • DAF channel reduction ratio r = 4
    Reduction ratio of the channel attention MLP in the Global Stream Attention branch, 'set to 4 by default'; a standard CBAM-style choice.
  • Local window sampling ratio = 0.5
    Only half of the local windows are randomly selected for SSM scanning per pass (Eq. 9); the ratio is chosen by hand and doubles as a dropout-like regularizer.
assumptions (4)
  • standard math The ZOH discretization formulas in Eq. 2 correctly describe the Mamba-style SSM implementation used in both branches.
    Invoked in the Preliminaries section; standard material from Gu and Dao 2023, stated without proof.
  • domain assumption RSCD's 27-class labels (friction, material, and unevenness combinations) are accurate, and the train/val split is used as intended.
    Invoked in the Datasets section; the paper never states which split of RSCD it uses, how many training images, or how class imbalance is handled.
  • domain assumption Road surface semantics split cleanly into global (material, friction) and local (unevenness, damage) visual information.
    Stated in the Introduction as the motivation for the dual-branch design; if false, the task-specific advantage of the architecture disappears.
  • ad hoc to paper Randomly zeroing half of the local windows each pass preserves enough local signal for the local branch to learn useful texture features.
    Eq. 9 and the surrounding discussion; this design choice is specific to this paper and has no independent support beyond the reported ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RoadMamba: A Dual Branch Visual State Space Model for Road Surface Classification." pith.science (2026). https://pith.science/paper/547YH2LF

@misc{pith2026250801210,
  author       = {Pith},
  title        = {Pith review of: RoadMamba: A Dual Branch Visual State Space Model for Road Surface Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/547YH2LF}},
  note         = {Machine review of arXiv:2508.01210}
}
read the original abstract

Acquiring the road surface conditions in advance based on visual technologies provides effective information for the planning and control system of autonomous vehicles, thus improving the safety and driving comfort of the vehicles. Recently, the Mamba architecture based on state-space models has shown remarkable performance in visual processing tasks, benefiting from the efficient global receptive field. However, existing Mamba architectures struggle to achieve state-of-the-art visual road surface classification due to their lack of effective extraction of the local texture of the road surface. In this paper, we explore for the first time the potential of visual Mamba architectures for road surface classification task and propose a method that effectively combines local and global perception, called RoadMamba. Specifically, we utilize the Dual State Space Model (DualSSM) to effectively extract the global semantics and local texture of the road surface and decode and fuse the dual features through the Dual Attention Fusion (DAF). In addition, we propose a dual auxiliary loss to explicitly constrain dual branches, preventing the network from relying only on global semantic information from the deep large receptive field and ignoring the local texture. The proposed RoadMamba achieves the state-of-the-art performance in experiments on a large-scale road surface classification dataset containing 1 million samples.

Figures

Figures reproduced from arXiv: 2508.01210 by the authors.

Figure 1
Figure 1. Architectural details of the RoadMamba. The model processes the input image through a stem and four stages of DualSSM Blocks, each containing a Local/Global SSM for parallel extraction of global and local features and a Dual Attention Fusion module to enhance extracted local and global features with auxiliary losses. Multi-scale features from all stages are concatenated and fed into the classification head for final… view at source ↗
Figure 3
Figure 3. present the structural details of the Dual Attention Fusion (DAF) unit within the DualSSM module. The DAF module is designed to adaptively fuse features extracted from both the global and local branches, thereby fully lever￾aging the strengths of each. This unit comprises two com￾plementary attention branches: Global Stream Attention and Local Stream Attention, which operate on global and local features, respectivel… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Defer to Plan: Adaptive Multi-Agent Fusion for End-to-End V2X Driving

    cs.RO 2026-07 conditional novelty 6.0 of 10

    Moving multi-agent fusion from perception to planning, via an autoregressive decoder with MoE tokenization, yields 79.72 driving score on V2Xverse vs CoDriving's 77.15.

Reference graph

Works this paper leans on

31 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    H.; Vora, S.; Liong, V

    Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; and Beijbom, O. 2020. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11621--11631

    2020

  4. [4]

    Chen, Y.; Qin, H.; Zhang, Z.; Magno, M.; Benini, L.; and Li, Y. 2025. Q-mambair: Accurate quantized mamba for efficient image restoration. arXiv preprint arXiv:2503.21970

    arXiv 2025

  5. [5]

    Dhiman, A.; and Klette, R. 2019. Pothole detection using computer vision and learning. IEEE Transactions on Intelligent Transportation Systems, 21(8): 3536--3550

    work page 2019

  6. [6]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

    arXiv 2020

  7. [7]

    Geiger, A.; Lenz, P.; Stiller, C.; and Urtasun, R. 2013. Vision meets robotics: The kitti dataset. The international journal of robotics research, 32(11): 1231--1237

    work page 2013

  8. [8]

    Gu, A.; and Dao, T. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

    arXiv 2023

Show all 31 references
  1. [9]

    Gu, A.; Goel, K.; and R \'e , C. 2021. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396

    2021 arXiv

  2. [10]

    Guo, H.; Li, J.; Dai, T.; Ouyang, Z.; Ren, X.; and Xia, S.-T. 2024. Mambair: A simple baseline for image restoration with state-space model. In European conference on computer vision, 222--241. Springer

    2024

  3. [11]

    Hatamizadeh, A.; and Kautz, J. 2025. Mambavision: A hybrid mamba-transformer vision backbone. In Proceedings of the Computer Vision and Pattern Recognition Conference, 25261--25270

    2025

  4. [12]

    Huang, T.; Pei, X.; You, S.; Wang, F.; Qian, C.; and Xu, C. 2024. Localmamba: Visual state space model with windowed selective scan. In European Conference on Computer Vision, 12--22. Springer

    2024

  5. [13]

    Kalman, R. E. 1960. A new approach to linear filtering and prediction problems

    1960

  6. [14]

    Leng, B.; Jin, D.; Xiong, L.; Yang, X.; and Yu, Z. 2021. Estimation of tire-road peak adhesion coefficient for intelligent electric vehicles based on camera and tire dynamics information fusion. Mechanical Systems and Signal Processing, 150: 107275

    2021

  7. [15]

    Liu, Y.; Tian, Y.; Zhao, Y.; Yu, H.; Xie, L.; Wang, Y.; Ye, Q.; Jiao, J.; and Liu, Y. 2024. Vmamba: Visual state space model. Advances in neural information processing systems, 37: 103031--103063

    2024

  8. [16]

    Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, 10012--10022

    2021

  9. [17]

    Liu, Z.; Mao, H.; Wu, C.-Y.; Feichtenhofer, C.; Darrell, T.; and Xie, S. 2022. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11976--11986

    2022

  10. [18]

    Nolte, M.; Kister, N.; and Maurer, M. 2018. Assessment of deep convolutional neural networks for road surface classification. In 2018 21st International Conference on Intelligent Transportation Systems (ITSC), 381--386. IEEE

    2018

  11. [19]

    Roychowdhury, S.; Zhao, M.; Wallin, A.; Ohlsson, N.; and Jonasson, M. 2018. Machine learning models for road surface and friction estimation using front-camera images. In 2018 International Joint Conference on Neural Networks (IJCNN), 1--8. IEEE

    2018

  12. [20]

    S abanovi c , E.; Z uraulis, V.; Prentkovskis, O.; and Skrickij, V. 2020. Identification of road-surface type using deep neural networks for friction coefficient estimation. Sensors, 20(3): 612

    2020

  13. [21]

    T.; Warrington, A.; and Linderman, S

    Smith, J. T.; Warrington, A.; and Linderman, S. W. 2022. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933

    2022 arXiv

  14. [22]

    Sun, P.; Kretzschmar, H.; Dotiwalla, X.; Chouard, A.; Patnaik, V.; Tsui, P.; Guo, J.; Zhou, Y.; Chai, Y.; Caine, B.; et al. 2020. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern reco...

    2020

  15. [23]

    Tian, C.; Jin, D.; Leng, B.; and Xiong, L. 2021. Reliable identification of road surface condition considering shadow interference. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), 251--257. IEEE

    2021

  16. [24]

    Wang, T.; Zhang, Z.; and Sun, C. 2025. RoadFormer: Local-Global Feature Fusion for Road Surface Classification in Autonomous Driving. arXiv preprint arXiv:2506.02358

    2025 arXiv

  17. [25]

    Yang, L.; Yu, K.; Tang, T.; Li, J.; Yuan, K.; Wang, L.; Zhang, X.; and Chen, P. 2023. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21611--21620

    2023

  18. [26]

    Yu, H.; Luo, Y.; Shu, M.; Huo, Y.; Yang, Z.; Shi, Y.; Guo, Z.; Li, H.; Hu, X.; Yuan, J.; et al. 2022. Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

    2022

  19. [27]

    Zhang, Z.; Sun, C.; Wang, B.; Guo, B.; Wen, D.; Zhu, T.; and Ning, Q. 2025 a . Height3d: A roadside visual framework based on height prediction in real 3-d space. IEEE Transactions on Intelligent Transportation Systems

    2025

  20. [28]

    Zhang, Z.; Sun, C.; Yue, C.; Wen, D.; Chen, Y.; Wang, T.; and Leng, J. 2025 b . Heightformer: Learning height prediction in voxel features for roadside vision centric 3d object detection via transformer. arXiv preprint arXiv:2503.10777

    2025 arXiv

  21. [29]

    Zhang, Z.; Sun, C.; Yue, C.; Wen, D.; Wang, T.; and Leng, J. 2025 c . Pillarmamba: Learning local-global context for roadside point cloud via hybrid state space model. arXiv preprint arXiv:2505.05397

    2025 arXiv

  22. [30]

    Zhao, T.; He, J.; Lv, J.; Min, D.; and Wei, Y. 2023. A comprehensive implementation of road surface classification for vehicle driving assistance: Dataset, models, and deployment. IEEE Transactions on Intelligent Transportation Systems, 24(8): 8361--8370

    2023

  23. [31]

    Zhu, L.; Liao, B.; Zhang, Q.; Wang, X.; Liu, W.; and Wang, X. 2024. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417

    2024 arXiv

Pith tools

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