Pith. sign in

REVIEW 4 major objections 5 minor 46 references

SRMambaV2: Biomimetic Attention for Sparse Point Cloud Upsampling in Autonomous Driving

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 A biomimetic range-image upsampler claims large gains in sparse LiDAR point cloud reconstruction for autonomous driving.

desk verdict A plausible incremental gain in range-image LiDAR upsampling, but the paper's own training-protocol ambiguity and missing error bars make the headline numbers hard to trust as written. read the letter →

arxiv 2507.17479 v1 pith:KVZQPTYC submitted 2025-07-23 cs.CV cs.LG

Chuang Chen , Xiaolin Qin , Jing Hu , Wenyi Ge This is my paper
classification cs.CVcs.LG
keywords LiDARpointcloudupsamplingrangeimagesuper-resolutionselectivescanningstatespacemodelself-attentionautonomousdrivingsparsebird's-eye-viewconstraintprogressiveadaptiveloss
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that LiDAR point cloud upsampling can be improved by treating the 3D problem as a 2D range-image super-resolution task, and by borrowing a visual-attention strategy from human drivers: first scan the whole image globally, then focus on sparse distant regions. The proposed network, SRMambaV2, uses a 2D selective scanning self-attention module to build a global overview and then a shifted-window Transformer to refine it, guided by a 'scan-to-focus' prior. A progressive adaptive loss with depth weighting and a bird's-eye-view consistency term pushes the network to preserve geometry in distant, sparsely sampled areas. On KITTI-360 and nuScenes, the authors report consistent improvements over the strongest published baseline TULIP across IoU, Chamfer distance, mean absolute error, and Jensen-Shannon divergence.

What carries the argument

The load-bearing object is the 2D Selective Scanning Self-Attention (2DSSA) module, which implements a three-stage 'scan-to-focus' pipeline. In the scanning stage, an SS2D block sweeps the feature map in four directions (left-to-right, right-to-left, top-to-bottom, bottom-to-top) to form a global overview; the modulation stage then recalibrates channels with a modified squeeze-and-excitation layer and a dual down/upsampling branch; the focus stage feeds the result as a context prior into a Swin Transformer with window and shifted-window attention. The second load-bearing element is the progressive adaptive loss (PAL), which adds a learnable region-weight mask to the usual L1 loss and a bird's-eye-view consistency term computed in projected top-down space, so the network is penalised for producing points that push geometry off the ground plane.

What would settle it

Run the same training and evaluation protocol on native low-beam LiDAR data (e.g., raw 8- or 16-beam sweeps with their real scan patterns and noise) rather than uniformly subsampled dense scans, and measure depth-stratified IoU and Chamfer distance; if the advantage over TULIP shrinks or disappears in far-field bins, the claim that the method generalises to real sparse sensors is unsubstantiated.

Watch

Extended reading notes

Core claim

SRMambaV2 establishes that a Mamba-style selective scan combined with windowed self-attention, arranged in a scan-to-focus sequence, yields higher-fidelity range-image super-resolution for LiDAR than either pure attention (TULIP) or pure selective-scan (SRMamba) baselines. The gain is concentrated at long range: depth-stratified analysis on KITTI-360 and nuScenes shows the method keeps IoU and Chamfer distance better than baselines in far-field bins, where point density is lowest. The paper's headline numbers on KITTI-360 are IoU improving from 0.4152 (TULIP) to 0.4516, Chamfer distance from 0.1241 to 0.0826, MAE from 0.0051 to 0.0041, and JSD from 0.0070 to 0.0045.

Load-bearing premise

The evaluation simulates sparse LiDAR by uniformly removing three out of every four points from dense scans, whereas real low-beam LiDAR sensors have different beam elevation patterns, noise, and occlusion, so the reported long-range gains may not replicate on actual 8-beam or 16-beam hardware.

Editorial extensions

If this is right

  • If the gains hold, range-image-based upsampling becomes a more viable route for densifying low-cost LiDAR in autonomous driving, since the method improves far-field perception without extra sensor hardware.
  • The scan-to-focus design suggests that Mamba-style global scanning can serve as an effective prior for Transformer attention on sparse depth imagery, not just RGB images.
  • The PAL loss shows that supervising in both range-image space and bird's-eye-view projected space reduces discrete noise points, which should improve downstream tasks that rely on clean ground and object contours.
  • Because the method targets the vertical (beam) dimension only, it could be combined with any horizontal-only densification or temporal fusion approach to push effective resolution further.
  • Consistent improvements across both KITTI-360 and nuScenes, two different sensors and dataset scales, indicate the method is not tuned to a single beam configuration.

Reading between the lines

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

  • The depth-stratified results suggest the real signal is in far-field bins; a natural stress test is to plot gain versus range and see if the margin grows monotonically, and to check whether the method still helps when the training data excludes near-range points.
  • Because the paper compares against TULIP and SRMamba at matched depth [2,2,2,2], the relative gain may shrink at larger capacities; testing at deeper configurations would clarify whether scan-to-focus is a scaling advantage or just a better initialisation.
  • The uniform-4x subsampling assumption is the fragile link: real 16-beam sensors have non-uniform vertical angles and systematic dropouts from reflectivity, so an evaluation on native low-beam sweeps with the same voxelised metrics would directly probe the claimed practical value.
  • If the BEV consistency constraint is truly what suppresses ghost points, an ablation that removes only that term should show a spike in Chamfer distance and isolated-point noise; the paper's grouped ablation mixes 2DSSA and modulation, so the individual contribution is still open.
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 SRMambaV2, a range-image-based point cloud upsampling method for autonomous driving LiDAR. The method combines a 2D selective scanning self-attention (2DSSA) mechanism, a dual-branch U-Net architecture with scan-to-focus (S2F) stages, and a progressive adaptive loss (PAL) that combines L1 loss, a learned region-weighting mask, and a bird's-eye-view (BEV) consistency term. Experiments on KITTI-360 and nuScenes report state-of-the-art results against image super-resolution methods and prior LiDAR upsampling baselines, including TULIP and the authors' earlier SRMamba. The paper also presents qualitative visualizations, depth-stratified metrics, an ablation study, and a failure case.

Significance. If the reported results are reliable, the paper offers a practically relevant contribution to LiDAR point cloud upsampling: the biomimetic S2F design is intuitively motivated, and the PAL loss with BEV supervision is a sensible way to address range-image-specific artifacts. The qualitative visualizations suggest genuine improvements in sparse distant regions. However, the manuscript's current evaluation has several gaps that prevent verification of the central SOTA claim: the training-protocol inconsistency between Sections IV-A.3 and IV-C, missing metric definitions and error bars, and an unexplained anomalous MAN CD value. These issues are fixable in a revision, and the core ideas appear defensible, so the paper merits a major revision rather than rejection.

major comments (4)
  1. [IV-A.3 and IV-C/Fig. 10] The training protocol is internally inconsistent. Section IV-A.3 states that models are trained for 1000 epochs with PAL as part of the total loss from the start, but Section IV-C and Figure 10 describe 1300 epochs with the loss function introduced at epoch 950. It is never stated whether Table I and Table II numbers come from the 1000-epoch from-start protocol, the 1300-epoch late-introduction protocol, or a mixture. This is load-bearing because the PAL ablation row in Table II cannot be interpreted correctly unless the schedule is specified; if PAL only helps when activated after L1 pretraining, then the from-start comparison does not isolate the loss contribution. Please state the exact protocol for each reported number and report the values of alpha and beta in Eq. (4).
  2. [Table I and Section IV-B.2] The quantitative evaluation lacks essential detail. The Chamfer Distance formula and the voxel resolution used for IoU are not defined, and no error bars or multiple-run statistics are provided. Given that the claimed improvements over SRMamba are small (e.g., MAE 0.0041 vs. 0.0044 on KITTI-360), single-run numbers cannot support the strength of the claims. Additionally, the MAN CD value of 1.1258 on KITTI-360 is an order of magnitude larger than all other methods (TULIP 0.1241, SRMamba 0.1031) despite MAN's MAE being comparable to TULIP's; this inconsistency suggests either a typo or a different CD formulation, and it must be corrected or explained.
  3. [IV-A.3 and Section IV-B] The evaluation setup does not support claims of practical value for autonomous driving. The sparse input is simulated by downsampling dense scans by a factor of four, but the downsampling procedure is not described, and the reported results are obtained on the same validation set used for model selection. There is no separate held-out test set, so the reported margins may reflect validation-set overfitting. Please specify the downsampling method (e.g., uniform random per-scan, beam subsampling), add a disjoint test split or cross-validation, and ideally evaluate on real low-beam LiDAR data or at least discuss the expected domain gap.
  4. [Table II and Section IV-C] The ablation study's baseline row is numerically identical to the SRMamba [6] row in Table I (IoU 0.4389, CD 0.1031, MAE 0.0044 on KITTI-360). Please clarify whether BL is SRMamba. If so, the paper should state this explicitly, because the ablation then measures the incremental contribution of 2DSSA and PAL over the authors' own prior method; this is acceptable but must be transparent. If BL is a different configuration, the coincidence of numbers needs explanation. Also, the 'progressive' aspect of PAL is never formally defined; the only progressive element described is the delayed introduction at epoch 950, which conflicts with the Section IV-A.3 protocol.
minor comments (5)
  1. [I (Introduction)] The acronym 'SRMamb' appears in the sentence 'We also observe that SRMamb [6] is capable of effectively extracting object boundary contour information'; it should be 'SRMamba'.
  2. [IV-B.2] The sentence 'it achieves an order-of-magnitude reduction in the CD' is not supported by the comparison against TULIP (33.4% reduction, about 1.5x) in the same paragraph; if the claim refers to a different baseline (e.g., Bilinear), that should be stated explicitly.
  3. [Appendix A] The heading reads 'Appendx A' and should be 'Appendix A'. In Eq. (7), the expression 'arctan(∆b − zi, sqrt(x_i^2 + y_i^2))' uses a two-argument arctangent without defining it; use 'atan2' notation for clarity.
  4. [III-C] The BEV constraint in Eq. (6) depends on grid resolutions H_bev and W_bev, but these values are never reported; please provide them or state how they are determined.
  5. [References] Reference [6] is listed as '2025' with no venue and appears likely to be the authors' own prior work; its SRMamba results in Table I should be independently reproducible, and the reference should be completed with a venue or arXiv identifier if available.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the SOTA claim is an empirical benchmark result against external baselines on held-out validation, not a derivation from its inputs.

full rationale

SRMambaV2's central claim (Sec. IV-B.2, Table I) is that it outperforms TULIP [5] and other baselines on KITTI-360 and nuScenes. That claim is established by training on the training split and evaluating IoU, CD, MAE, and JSD on held-out validation scans, so it is not a derivation that reduces to its inputs. The loss (Eq. 4) and network components are fitted by standard supervised training, and no reported metric is defined in terms of a fitted parameter or a self-referential equation. The only related-paper concern is that the SRMamba baseline (ref [6]) is the authors' own prior work and its Table I numbers may come from the same training pipeline rather than from the original publication's independently reported values; however, the primary SOTA comparison is against the external TULIP [5], and the improvement over SRMamba is a measured difference on the same held-out data, not a circular argument. The inconsistency between the 1000-epoch protocol in Sec. IV-A.3 and the 1300-epoch/PAL-at-epoch-950 experiment in Sec. IV-C and Fig. 10 is a reproducibility and attribution problem, not circularity. The paper also discloses its own failure case in Sec. IV-D, which further confirms the evaluation is empirical. No equation or citation makes the output equivalent to the input by construction, so no load-bearing circular step is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a network with many trainable weights, which are not counted here. The unstated loss weights alpha and beta and the unspecified BEV grid resolution are the main manual free parameters; the subsampling-based data generation is a domain assumption that could break the practical claim. No new physical entities are introduced.

free parameters (5)
  • alpha (BEV loss weight)
    In Eq. (4), alpha scales the bird's-eye-view constraint. The paper never reports its value, only that it controls relative importance.
  • beta (adaptive L1 loss weight)
    In Eq. (4), beta scales the adaptive L1 term. Value not reported in Sec. III-C or IV.
  • Swin patch size = 2x8
    Chosen in Sec. III-B.3 to match the 1:8 vertical-to-horizontal aspect ratio of range images; a hand-picked design choice.
  • Network depth = [2,2,2,2]
    Fixed for fair comparison across methods (Sec. IV-B.1); no sensitivity study.
  • BEV grid resolution (Hbev x Wbev)
    Eq. (6) defines Lbev over Hbev x Wbev, but the paper states these dimensions are 'determined by the distribution of point clouds' without giving the actual values.
assumptions (5)
  • standard math Spherical projection and back-projection formulas (Appendix A) correctly map 3D coordinates to range image pixels and back.
    Used in Sec. III-A to define the task and in the BEV loss; standard sensor geometry, not derived in this paper.
  • domain assumption Distant sparse areas of the range image lie in the upper rows and contain concentrated, structurally critical features.
    Sec. I: 'points corresponding to distant areas in range images are usually located in the upper rows of the image, exhibiting a sparse yet relatively concentrated spatial feature distribution.' This motivates the loss weighting.
  • domain assumption Human coarse-to-fine visual attention is an appropriate design template for range-image feature extraction.
    Sec. I and Sec. III-B frame the scan-to-focus strategy as biomimetic; this analogy is motivational, not experimentally validated.
  • ad hoc to paper Back-projecting upsampled range images and applying an L1 loss in bird's-eye-view space provides useful geometric supervision.
    The BEV constraint in Eq. (6) is introduced to reduce spatial offsets (Fig. 5b); its benefit is shown only in the final ablation row, not independently.
  • domain assumption Uniformly subsampling dense scans by 4x produces training inputs representative of real low-beam LiDAR.
    Sec. IV-A.3 uses this to construct sparse inputs; the realism of this simulation is not verified against real 8- or 16-beam sensors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SRMambaV2: Biomimetic Attention for Sparse Point Cloud Upsampling in Autonomous Driving." pith.science (2026). https://pith.science/paper/KVZQPTYC

@misc{pith2026250717479,
  author       = {Pith},
  title        = {Pith review of: SRMambaV2: Biomimetic Attention for Sparse Point Cloud Upsampling in Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KVZQPTYC}},
  note         = {Machine review of arXiv:2507.17479}
}
read the original abstract

Upsampling LiDAR point clouds in autonomous driving scenarios remains a significant challenge due to the inherent sparsity and complex 3D structures of the data. Recent studies have attempted to address this problem by converting the complex 3D spatial scenes into 2D image super-resolution tasks. However, due to the sparse and blurry feature representation of range images, accurately reconstructing detailed and complex spatial topologies remains a major difficulty. To tackle this, we propose a novel sparse point cloud upsampling method named SRMambaV2, which enhances the upsampling accuracy in long-range sparse regions while preserving the overall geometric reconstruction quality. Specifically, inspired by human driver visual perception, we design a biomimetic 2D selective scanning self-attention (2DSSA) mechanism to model the feature distribution in distant sparse areas. Meanwhile, we introduce a dual-branch network architecture to enhance the representation of sparse features. In addition, we introduce a progressive adaptive loss (PAL) function to further refine the reconstruction of fine-grained details during the upsampling process. Experimental results demonstrate that SRMambaV2 achieves superior performance in both qualitative and quantitative evaluations, highlighting its effectiveness and practical value in automotive sparse point cloud upsampling tasks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 44 canonical work pages

  1. [6]

    Srmamba: Mamba for super-resolution of lidar point clouds,

    C. Chen and W. Ge, “Srmamba: Mamba for super-resolution of lidar point clouds,” 2025

    work page 2025

  2. [1]

    Efficient deep super-resolution of voxelized point cloud in geometry compression,

    K. Matsuzaki and S. Komorita, “Efficient deep super-resolution of voxelized point cloud in geometry compression,” IEEE Sensors Journal , vol. 23, no. 2, pp. 1328–1342, 2023

    work page 2023

  3. [2]

    Farvnet: A fast and accurate range-view-based method for semantic segmentation of point clouds,

    C. Chen, L. Zhao, W. Guo, X. Yuan, S. Tan, J. Hu, Z. Yang, S. Wang, and W. Ge, “Farvnet: A fast and accurate range-view-based method for semantic segmentation of point clouds,” Sensors, vol. 25, no. 9, 2025

    work page 2025

  4. [3]

    Subt-mrs dataset: Pushing slam towards all-weather environments,

    S. Zhao, Y . Gao, T. Wu, D. Singh, R. Jiang, H. Sun, M. Sarawata, Y . Qiu, W. Whittaker, I. Higgins, Y . Du, S. Su, C. Xu, J. Keller, J. Karhade, L. Nogueira, S. Saha, J. Zhang, W. Wang, C. Wang, and S. Scherer, “Subt-mrs dataset: Pushing slam towards all-weather environments,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ...

    work page 2024

  5. [4]

    A multi-source fusion system for through-wall radar compensation using lidar and slam-based 3d reconstruction,

    X. Zeng, Y . Hu, X. Yang, Z. Yin, and S. Zhong, “A multi-source fusion system for through-wall radar compensation using lidar and slam-based 3d reconstruction,” IEEE Sensors Journal , pp. 1–1, 2025

    work page 2025

  6. [5]

    Tulip: Transformer for upsampling of lidar point clouds,

    B. Yang, P. Pfreundschuh, R. Siegwart, M. Hutter, P. Moghadam, and V . Patil, “Tulip: Transformer for upsampling of lidar point clouds,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR) , pp. 15354–15364, 2024

    work page 2024

  7. [7]

    Rangeldm: Fast realistic lidar point cloud generation,

    Q. Hu, Z. Zhang, and W. Hu, “Rangeldm: Fast realistic lidar point cloud generation,” in Computer Vision – ECCV 2024 (A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, eds.), (Cham), pp. 115–135, Springer Nature Switzerland, 2025

    work page 2024

  8. [8]

    Simulation-based lidar super-resolution for ground vehicles,

    T. Shan, J. Wang, F. Chen, P. Szenher, and B. Englot, “Simulation-based lidar super-resolution for ground vehicles,” Robotics and Autonomous Systems, vol. 134, p. 103647, 2020

    work page 2020

Show all 46 references
  1. [9]

    Ligapu: A lidar point cloud upsampling network for multiple complex scenes,

    B. Fu, M. Sui, H. Li, F. Yang, H. Yao, T. Deng, and X. Zhang, “Ligapu: A lidar point cloud upsampling network for multiple complex scenes,” Pattern Recognition, p. 112132, 2025

    2025

  2. [10]

    Neural mecha- nisms of visual attention: How top-down feedback highlights relevant locations,

    Y . B. Saalmann, I. N. Pigarev, and T. R. Vidyasagar, “Neural mecha- nisms of visual attention: How top-down feedback highlights relevant locations,” Science, p. 1612–1615, Jun 2007

    2007

  3. [11]

    Brain states: Top-down influences in sensory processing,

    C. D. Gilbert and M. Sigman, “Brain states: Top-down influences in sensory processing,” Neuron, vol. 54, p. 677–696, Jun 2007

    2007

  4. [12]

    Smfanet: A lightweight self-modulation feature aggregation network for efficient image super- resolution,

    M. Zheng, L. Sun, J. Dong, and J. Pan, “Smfanet: A lightweight self-modulation feature aggregation network for efficient image super- resolution,” in Computer Vision – ECCV 2024 (A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, eds.), (Cham), pp. 359–...

    2024

  5. [13]

    Learning to combine top-down and bottom- up signals in recurrent neural networks with attention over modules,

    S. Mittal, A. Lamb, A. Goyal, V . V oleti, M. Shanahan, G. Lajoie, M. Mozer, and Y . Bengio, “Learning to combine top-down and bottom- up signals in recurrent neural networks with attention over modules,” in Proceedings of the 37th International Conference on Machine Learning ...

    2020

  6. [14]

    Blend- mask: Top-down meets bottom-up for instance segmentation,

    H. Chen, K. Sun, Z. Tian, C. Shen, Y . Huang, and Y . Yan, “Blend- mask: Top-down meets bottom-up for instance segmentation,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2020

    2020

  7. [15]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 770–778, 2016

    2016

  8. [16]

    Fully convolutional networks for semantic segmentation,

    E. Shelhamer, J. Long, and T. Darrell, “Fully convolutional networks for semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 4, pp. 640–651, 2017

    2017

  9. [17]

    Overlock: An overview-first-look-closely-next convnet with context-mixing dynamic kernels,

    M. Lou and Y . Yu, “Overlock: An overview-first-look-closely-next convnet with context-mixing dynamic kernels,” in Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) , pp. 128– 138, June 2025

    2025

  10. [18]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “Vmamba: Visual state space model,” Curran Associates, Inc. , vol. 37, pp. 103031–103063, 2024

    2024

  11. [19]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9992–10002, 2021

    2021

  12. [20]

    Point set surfaces,

    M. Alexa, J. Behr, D. Cohen-Or, S. Fleishman, D. Levin, and C. Silva, “Point set surfaces,” in Proceedings Visualization, 2001. VIS ’01. , pp. 21–29, 537, 2001

    2001

  13. [21]

    Robust moving least- squares fitting with sharp features,

    S. Fleishman, D. Cohen-Or, and C. T. Silva, “Robust moving least- squares fitting with sharp features,” ACM Trans. Graph. , vol. 24, p. 544–552, July 2005

    2005

  14. [22]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    R. Q. Charles, H. Su, M. Kaichun, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 77–85, 2017

    2017

  15. [23]

    Pointnet++: deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: deep hierarchical feature learning on point sets in a metric space,” in Proceedings of the 31st International Conference on Neural Information Processing Sys- tems, NIPS’17, (Red Hook, NY , USA), p. 5105–5114, Curran Assoc...

    2017

  16. [24]

    Pu-dense: Sparse tensor-based point cloud geometry upsampling,

    A. Akhtar, Z. Li, G. V . d. Auwera, L. Li, and J. Chen, “Pu-dense: Sparse tensor-based point cloud geometry upsampling,” IEEE Transactions on Image Processing, vol. 31, pp. 4133–4148, 2022

    2022

  17. [25]

    Pvt: An implicit surface reconstruc- tion framework via point voxel geometric-aware transformer,

    C. Fan, C. Zhao, and Y . Duan, “Pvt: An implicit surface reconstruc- tion framework via point voxel geometric-aware transformer,” in 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 3013–3023, 2025

    2025

  18. [26]

    A fast ground segmentation method of lidar point cloud from coarse-to-fine,

    D. Guo, G. Yang, B. Qi, and C. Wang, “A fast ground segmentation method of lidar point cloud from coarse-to-fine,” IEEE Sensors Journal , vol. 23, no. 2, pp. 1357–1367, 2023

    2023

  19. [27]

    Rangelvdet: Boosting 3d object detection in lidar with range image and rgb image,

    Z. Zhang, Z. Liang, M. Zhang, X. Zhao, H. Li, M. Yang, W. Tan, and S. Pu, “Rangelvdet: Boosting 3d object detection in lidar with range image and rgb image,” IEEE Sensors Journal , vol. 22, no. 2, pp. 1391– 1403, 2022. AUTHOR et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS...

    2022

  20. [28]

    Face mamba: A facial emotion analysis network based on vmamba*,

    L. Li, Q. Sun, L. Zhao, H. Sun, F. Zhao, and B. Gu, “Face mamba: A facial emotion analysis network based on vmamba*,” in 2024 7th International Conference on Machine Learning and Natural Language Processing (MLNLP), pp. 1–5, 2024

    2024

  21. [29]

    Hfifnet: Hierarchical feature interaction network with multiscale fusion for change detection,

    M. Han, T. Xu, Q. Liu, X. Yang, J. Wang, and J. Kong, “Hfifnet: Hierarchical feature interaction network with multiscale fusion for change detection,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 18, pp. 4318–4330, 2025

    2025

  22. [30]

    Intermamba: A visual-prompted interactive framework for dense object detection and annotation,

    S. Liu, Z. Yang, Q. Li, and Q. Wang, “Intermamba: A visual-prompted interactive framework for dense object detection and annotation,” IEEE Transactions on Geoscience and Remote Sensing , vol. 63, pp. 1–11, 2025

    2025

  23. [31]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 (N. Navab, J. Horneg- ger, W. M. Wells, and A. F. Frangi, eds.), (Cham), pp. 234–241, Springe...

    2015

  24. [32]

    Segumamba: Integrating mamba with u net for medical image segmentation,

    Y . Zhao, C. Liu, X. Zhou, and X. Zhang, “Segumamba: Integrating mamba with u net for medical image segmentation,” in 2024 Interna- tional Conference on Image Processing, Computer Vision and Machine Learning (ICICML) , pp. 108–111, 2024

    2024

  25. [33]

    Tmu: Transmission- enhanced mamba-unet for medical image segmentation,

    X. Yang, Z. Luo, Y . Wu, X. Xie, L. Nan, and T. Li, “Tmu: Transmission- enhanced mamba-unet for medical image segmentation,” in Advanced Intelligent Computing Technology and Applications (D.-S. Huang, C. Zhang, and J. Guo, eds.), (Singapore), pp. 428–438, Springer Nature Singa...

    2024

  26. [34]

    Lightmamba-unet: Lightweight mamba with u-net for efficient skin lesion segmentation,

    W. Hou, S. Zhou, and H. Zhao, “Lightmamba-unet: Lightweight mamba with u-net for efficient skin lesion segmentation,” in Advances in Brain Inspired Cognitive Systems (A. Hussain, B. Jiang, J. Ren, M. Mahmud, E. Yang, A. Zheng, C. Li, S. Wang, Z. Gao, and Z. Zhao, eds.), (Singa...

    2025

  27. [35]

    Mobilemamba: Lightweight multi-receptive visual mamba network,

    H. He, J. Zhang, Y . Cai, H. Chen, X. Hu, Z. Gan, Y . Wang, C. Wang, Y . Wu, and L. Xie, “Mobilemamba: Lightweight multi-receptive visual mamba network,” in Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) , pp. 4497–4507, June 2025

    2025

  28. [36]

    Localmamba: Visual state space model with windowed selective scan,

    T. Huang, X. Pei, S. You, F. Wang, C. Qian, and C. Xu, “Localmamba: Visual state space model with windowed selective scan,” in Computer Vision – ECCV 2024 Workshops (A. Del Bue, C. Canton, J. Pont-Tuset, and T. Tommasi, eds.), (Cham), pp. 12–22, Springer Nature Switzerland, 2025

    2024

  29. [37]

    Mambavision: A hybrid mamba- transformer vision backbone,

    A. Hatamizadeh and J. Kautz, “Mambavision: A hybrid mamba- transformer vision backbone,” in Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) , pp. 25261–25270, June 2025

    2025

  30. [38]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” in 2017 IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR) , pp. 936–944, 2017

    2017

  31. [39]

    Pyramid vision transformer: A versatile backbone for dense prediction without convolutions,

    W. Wang, E. Xie, X. Li, D.-P. Fan, K. Song, D. Liang, T. Lu, P. Luo, and L. Shao, “Pyramid vision transformer: A versatile backbone for dense prediction without convolutions,” in2021 IEEE/CVF International Conference on Computer Vision (ICCV) , pp. 548–558, 2021

    2021

  32. [40]

    An image is worth 16x16 words: Trans- formers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” in International Conference on Learni...

    2021

  33. [41]

    Multi-scale attention network for single image super-resolution,

    Y . Wang, Y . Li, G. Wang, and X. Liu, “Multi-scale attention network for single image super-resolution,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pp. 5950–5960, 2024

    2024

  34. [42]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun 2018

    2018

  35. [43]

    Kitti-360: A novel dataset and bench- marks for urban scene understanding in 2d and 3d,

    Y . Liao, J. Xie, and A. Geiger, “Kitti-360: A novel dataset and bench- marks for urban scene understanding in 2d and 3d,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 3, pp. 3292– 3310, 2023

    2023

  36. [44]

    Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking,

    W. K. Fong, R. Mohan, J. V . Hurtado, L. Zhou, H. Caesar, O. Beijbom, and A. Valada, “Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking,” IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 3795–3802, 2022

    2022

  37. [45]

    Cas-vit: Convolutional additive self-attention vision transformers for efficient mobile applications,

    T. Zhang, L. Li, Y . Zhou, W. Liu, C. Qian, and X. Ji, “Cas-vit: Convolutional additive self-attention vision transformers for efficient mobile applications,” arXiv preprint arXiv:2408.03703 , 2024

    2024 arXiv

  38. [46]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” in 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) , pp. 1833–1844, 2021

    2021

Pith tools

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