pith. sign in

arxiv: 2605.15546 · v1 · pith:KAT6LIHWnew · submitted 2026-05-15 · 💻 cs.CV

3DTMDet: A Dual-Path Synergy Network of Transformer and SSM for 3D Object Detection in Point Clouds

Bingwen Qiu , Yuan Liu , Junqi Bai , Tong Jiang , Ben Liang , Fangzhou Chen , Xiubao Sui , Qian Chen This is my paper

Pith reviewed 2026-05-19 14:15 UTC · model grok-4.3

classification 💻 cs.CV
keywords 3D object detectionpoint cloudsstate space modelTransformerhybrid networkKITTIONCE
0
0 comments X p. Extension
pith:KAT6LIHW Add to your LaTeX paper What is a Pith Number? 
\usepackage{pith}
\pithnumber{KAT6LIHW}

Prints a linked pith:KAT6LIHW badge after your title and writes the identifier into PDF metadata. Compiles on arXiv with no extra files. Learn more

The pith

A hybrid Transformer and state space model network better detects objects in sparse distant point clouds.

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

The core problem is the tension between needing wide context for remote points and keeping local geometric details that are scarce in distant sparse data. Standard 1D serialization methods enlarge the field of view but lose those details. 3DTMDet solves this by pairing Mamba state space models for efficient long sequences with Transformer attention for local structures. Its 3D Hybrid Mamba Transformer block runs features through an SSM-Attention-SSM sequence to blend global and local processing. A voxel generation step spreads features along the sensor line of sight to fill occluded regions, leading to better results on KITTI and ONCE than prior detectors.

Core claim

By integrating state space models with Transformers through an SSM-Attention-SSM pipeline in the 3D Hybrid Mamba Transformer block and adding a physics-inspired voxel generation block, the network achieves superior 3D object detection by maintaining local geometric information while capturing global context in sparse point clouds.

What carries the argument

The 3D Hybrid Mamba Transformer block that applies state space model, attention, and state space model in sequence to balance global and local feature processing.

If this is right

  • Improved accuracy in detecting small and distant objects due to preserved local details.
  • Outperformance of existing detectors on the KITTI and ONCE benchmarks.
  • Effective handling of occlusions through feature diffusion along observation directions.
  • Linear complexity global modeling that scales better than pure attention mechanisms for large point sets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This suggests hybrid architectures may be useful for other tasks with sparse 3D data such as scene reconstruction.
  • Testing the method on additional datasets or real-world LiDAR streams could validate broader applicability.
  • The approach opens questions about optimal ordering in SSM-attention hybrids for different data densities.

Load-bearing premise

The SSM-Attention-SSM pipeline balances global context understanding with preservation of fine-grained local geometric structures in sparse distant point sets.

What would settle it

Demonstrating no improvement in distant object detection precision on the ONCE dataset relative to prior methods would contradict the effectiveness of the proposed synergy.

Figures

Figures reproduced from arXiv: 2605.15546 by Ben Liang, Bingwen Qiu, Fangzhou Chen, Junqi Bai, Qian Chen, Tong Jiang, Xiubao Sui, Yuan Liu.

Figure 1
Figure 1. Figure 1: Illustration of the damage to the geometric structure caused by space-filling curves and windowing schemes. Both of these schemes may lead to the division of the unified target into different positions in the sequence, or even different sequences. This can lead to potential instability in linear operators. sequence, and proposes the establishment of a hierarchical structure using dual-scale SSM blocks, ena… view at source ↗
Figure 2
Figure 2. Figure 2: Top: The overall architecture of our proposed 3DTMDet. First, the point cloud data is converted into voxels and fed into a 3D backbone. The 3D backbone consists of multiple 3D Hybrid Mamba Transformer (3DHMT) Blocks, with voxels being compressed along Z-axis between modules. Internally, the 3DHMT Block consists of two Serialized-Mamba Blocks, one Grouped-Transformer Block and one Voxel Generation Block. Th… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the serialization scheme in Serialized￾Mamba Block. All voxels are sorted into equally sized groups along the X or Y axis, and arranged from low to high along the Z-axis. Each adjacent sets share the equal-sized overlapping voxels. data. Furthermore, we utilize absolute position encoding to emphasize global characteristics within this module, defined as follows: 𝐸𝑖 = 𝑀𝐿𝑃 (2𝜋𝑐𝑖 ) (1) where 𝑀… view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of our relative position encoding. Each voxel first computes the difference with its K-nearest neigh￾boring voxels to obtain the relative position. Then we directly encode the relative position through a MLP and then fused with main features [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustrations of our voxel generation scheme. The voxels are diffused away from the sensor, and the top-P diffused voxels are selected as the generated voxel. where, 𝑄𝐻 , 𝐾𝐻 and 𝑉𝐻 represent vectors from the Hilbert curve sequence, while 𝑄𝐻𝑇 , 𝐾𝐻𝑇 and 𝑉𝐻𝑇 represent vec￾tors from the Trans-Hilbert curve sequence. The local neighborhood contains key features of the point cloud data. The Grouped-Transformer b… view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of the basic sparse-conv block. This module contains two 3D Subm Sparse Convolutional layers and a residual link. This work replaces the Transformer Blocks or Mamba Blocks with this module to verify the effectiveness of our proposed network structure. local 3D geometric modeling, we achieve a robust balance that resolves the local-detail blindness of SSMs while miti￾gating the computational co… view at source ↗
Figure 8
Figure 8. Figure 8: Heatmap of voxel generation block conv for diffusion and eliminating the Top P Select step in our strategy. The experiment shows that our algorithm outperforms these two strategies by 0.7mAP and 0.5mAP, respectively. This indicates that selective diffusion is an effective means of enhancing the effectiveness of angle measurement. In this section, we also present visualizations of the heatmap of the generat… view at source ↗
Figure 9
Figure 9. Figure 9: The Qualitative Results Visualization of DSVT, LION and Ours. targets. This proves that for small objects such as pedes￾trians and cyclists, or for targets with long distances, the geometric information provided by point clouds is limited, and our strategy can better ensure detection performance. In Fig.9(d), our algorithm, compared to the pure Mamba algorithm Lion, missed detecting Car targets, but still … view at source ↗
read the original abstract

A fundamental challenge in point cloud object detection lies in the conflict between the extreme sparsity of distant points and the need for remote context understanding. The existing methods typically use 1D serialization to expand the receptive field, which inevitably discards already scarce local geometric details and reduces detection of distant and small objects. To address this issue, we propose 3DTMDet, a novel detection network that synergistically combines state space models (Mamba) with Transformers. The core idea is to utilize SSM's linear complexity and advantages in long sequence modeling to effectively capture global interactions between sparse and distant points, while using Transformer modules with local attention to encode fine-grained geometric structures in local point sets, preserving accurate shape information. We propose the 3D Hybrid Mamba Transformer (3DHMT) block, which uses an SSM-Attention-SSM pipeline to balance global context understanding and local detail preservation, effectively alleviating the tension between receptive field enlargement and geometric preservation in remote detection. In addition, we introduced a voxel generation block inspired by LiDAR physics, which diffuses features along the sensor observation direction to reconstruct the complete object structure of occlusion and distant areas. Extensive experiments conducted on the KITTI and ONCE datasets have shown that 3DTMDet outperforms state-of-the-art detectors. The code is available at https://github.com/QiuBingwen/3DTMDet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes 3DTMDet, a dual-path synergy network combining Transformers and state space models (Mamba/SSM) for 3D object detection in point clouds. It introduces the 3D Hybrid Mamba Transformer (3DHMT) block with an SSM-Attention-SSM pipeline to capture global context among sparse distant points while preserving local geometric details via attention, along with a voxel generation block that diffuses features along the LiDAR observation direction to reconstruct occluded or distant structures. The central claim is that this architecture outperforms state-of-the-art detectors on the KITTI and ONCE datasets.

Significance. If the empirical gains are shown to be robust and isolated to the proposed components, the work could meaningfully advance point-cloud detection by mitigating the receptive-field versus local-detail trade-off that limits performance on small and far objects. The hybrid SSM-Transformer design leverages Mamba's linear scaling for long-range modeling in a domain where quadratic attention is costly, and the physics-inspired voxel block offers a targeted way to handle sparsity.

major comments (3)
  1. [§4 Experiments] §4 (Experiments): The abstract states that 'extensive experiments... have shown that 3DTMDet outperforms state-of-the-art detectors' on KITTI and ONCE, yet no quantitative mAP values, baseline comparisons, error bars, or ablation tables are referenced in the provided description. Without explicit numbers and controls, the headline claim cannot be evaluated and remains load-bearing for the paper's contribution.
  2. [§3.2 3DHMT block] §3.2 (3DHMT block description): The claim that the SSM-Attention-SSM pipeline 'effectively balance[s] global context understanding and local detail preservation' is presented as the solution to the sparsity-receptive-field tension, but no ablation isolating the attention module versus pure SSM or pure Transformer variants is described. This leaves the central architectural assumption untested and risks confounding the reported gains with other factors.
  3. [§4 Experiments] §4 (Training protocol): No statement is given that all compared baselines were re-implemented and re-trained under identical optimizer, augmentation, and voxelization settings using the authors' codebase. In point-cloud detection, protocol mismatches routinely produce 1-3% mAP differences; without this control the attribution of gains to the 3DHMT block or voxel generation remains ambiguous.
minor comments (2)
  1. [Abstract] Abstract: The phrase 'diffuses features along the sensor observation direction' is introduced without a supporting equation or diagram reference, making the voxel generation block's mechanics hard to reconstruct from text alone.
  2. [§3 Method] Notation: The paper introduces '3D Hybrid Mamba Transformer (3DHMT) block' and 'voxel generation block' as new entities; a short table summarizing their input/output tensors and complexity would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We are grateful to the referee for the thoughtful and constructive feedback on our manuscript. We have carefully considered each comment and provide detailed responses below. Where appropriate, we have revised the manuscript to address the concerns raised.

read point-by-point responses

  1. Referee: [§4 Experiments] §4 (Experiments): The abstract states that 'extensive experiments... have shown that 3DTMDet outperforms state-of-the-art detectors' on KITTI and ONCE, yet no quantitative mAP values, baseline comparisons, error bars, or ablation tables are referenced in the provided description. Without explicit numbers and controls, the headline claim cannot be evaluated and remains load-bearing for the paper's contribution.

    Authors: We appreciate this observation. The abstract is intended as a concise summary and typically does not include detailed numerical results to maintain readability. However, the full manuscript in Section 4 provides extensive quantitative evaluations, including mAP values on both KITTI and ONCE datasets, comparisons with state-of-the-art methods, and ablation studies. To strengthen the abstract, we will revise it to include key performance highlights, such as the mAP improvements achieved. Additionally, we will ensure that all experimental results are clearly cross-referenced in the text. revision: yes

  2. Referee: [§3.2 3DHMT block] §3.2 (3DHMT block description): The claim that the SSM-Attention-SSM pipeline 'effectively balance[s] global context understanding and local detail preservation' is presented as the solution to the sparsity-receptive-field tension, but no ablation isolating the attention module versus pure SSM or pure Transformer variants is described. This leaves the central architectural assumption untested and risks confounding the reported gains with other factors.

    Authors: We thank the referee for highlighting the need for more targeted ablations. In the original manuscript, we included ablations on the overall 3DHMT block and its components. To directly address this, we will add a new ablation study that compares the full hybrid SSM-Attention-SSM pipeline against pure SSM and pure Transformer variants within the same framework. This will help isolate the contribution of the attention module in balancing global and local features. The revised manuscript will include this analysis in Section 4. revision: yes

  3. Referee: [§4 Experiments] §4 (Training protocol): No statement is given that all compared baselines were re-implemented and re-trained under identical optimizer, augmentation, and voxelization settings using the authors' codebase. In point-cloud detection, protocol mismatches routinely produce 1-3% mAP differences; without this control the attribution of gains to the 3DHMT block or voxel generation remains ambiguous.

    Authors: We agree that ensuring identical training protocols is crucial for fair comparisons. All baselines were indeed re-implemented and re-trained using the same settings in our codebase, including optimizer, data augmentation, and voxelization parameters. We will add an explicit statement in Section 4 (Experiments) detailing the training protocol and confirming that all methods were evaluated under identical conditions to eliminate any ambiguity in attributing performance gains. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical architecture proposal validated on external benchmarks

full rationale

The paper introduces 3DTMDet as a novel network design combining SSM (Mamba) and Transformer modules in a 3DHMT block plus a physics-inspired voxel generation block. All load-bearing claims rest on experimental outperformance on the public KITTI and ONCE datasets rather than any closed-form derivation, uniqueness theorem, or parameter fit that is then re-labeled as a prediction. No equations are presented that reduce by construction to the inputs; the SSM-Attention-SSM pipeline is motivated by stated desiderata (global context vs. local geometry) and evaluated empirically. Self-citations, if present, are not load-bearing for the central result. The work is therefore self-contained against external benchmarks with no detectable circular steps.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 2 invented entities

Abstract-only review limits precise enumeration; the work relies on standard assumptions about neural network expressivity and empirical validation. New components like the 3DHMT block are introduced without independent evidence beyond the reported experiments.

free parameters (1)
  • network hyperparameters and training settings
    Typical learned weights and tuning choices in deep learning detectors; not enumerated in abstract.
axioms (2)
  • domain assumption State space models provide linear-complexity long-sequence modeling suitable for global point interactions
    Invoked to justify SSM use for distant sparse points.
  • domain assumption Local Transformer attention preserves geometric details better than serialization
    Core motivation for the dual-path design.
invented entities (2)
  • 3D Hybrid Mamba Transformer (3DHMT) block no independent evidence
    purpose: Balance global context and local geometry via SSM-Attention-SSM pipeline
    Newly proposed architectural unit.
  • voxel generation block no independent evidence
    purpose: Diffuse features along sensor direction to reconstruct occluded or distant objects
    Inspired by LiDAR physics but presented as a new module.

pith-pipeline@v0.9.0 · 5807 in / 1457 out tokens · 72402 ms · 2026-05-19T14:15:03.468349+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · 1 internal anchor

  1. [1]

    A survey on deep-learning-based lidar 3d object detection for autonomous driving

    Alaba, S.Y., Ball, J.E., 2022. A survey on deep-learning-based lidar 3d object detection for autonomous driving. Sensors (Basel, Switzerland) 22. URL:https://api.semanticscholar.org/CorpusID: 254429128

    work page 2022

  2. [2]

    Neurocomputing 659, 131737

    Ao, L., Wan, W., Ouyang, N., Li, J., Li, Q., Gong, M., Sheng, K., 2026.Svp:stratifiedverticalpriorsforlidar-based3dobjectdetection. Neurocomputing 659, 131737. URL:https://www.sciencedirect. com/science/article/pii/S0925231225024099,doi:https://doi.org/10. 1016/j.neucom.2025.131737

    work page arXiv 2026

  3. [3]

    Dstr: Dual scenes transformer for cross-modal fusion in 3d object detection

    Cai, H., Yin, D., Yu, F.R., Xiong, S., 2025. Dstr: Dual scenes transformer for cross-modal fusion in 3d object detection. 2025 IEEE/CVF Winter Conference on Applications of Computer Vi- sion (WACV) , 3064–3073URL:https://api.semanticscholar.org/ CorpusID:277218349

    work page 2025

  4. [4]

    Kptr:Key point transformer for lidar-based 3d object detection

    Cao,J.,Peng,Y.,Wei,H.,Mo,L.,Fan,L.,Wang,L.,2025. Kptr:Key point transformer for lidar-based 3d object detection. Measurement 242,115820. URL:https://www.sciencedirect.com/science/article/ pii/S0263224124017056, doi:https://doi.org/10.1016/j.measurement. 2024.115820

    work page doi:10.1016/j.measurement 2025

  5. [5]

    Eb-lg module for 3d point cloud classification and segmentation

    Chen, J., Zhang, Y., Ma, F., Tan, Z., 2023. Eb-lg module for 3d point cloud classification and segmentation. IEEE Robotics and AutomationLetters8,160–167. URL:https://api.semanticscholar. org/CorpusID:253791066

    work page 2023

  6. [7]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    Gu, A., Dao, T., 2023. Mamba: Linear-time sequence modeling with selective state spaces. ArXiv abs/2312.00752. URL:https: //api.semanticscholar.org/CorpusID:265551773

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [8]

    M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers

    Guan, T., Wang, J., Lan, S., Chandra, R., Wu, Z., Davis, L.S., Manocha, D., 2021. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. 2022 IEEE/CVF Winter Conference on Applications of Computer Vi- sion (WACV) , 2293–2303URL:https://api.semanticscholar.org/ CorpusID:233394223

    work page 2021

  8. [9]

    Mambatron: Efficient cross- modalpointcloudenhancementusingaggregateselectivestatespace modeling

    Inaganti, S.T., Petrenko, G., 2025. Mambatron: Efficient cross- modalpointcloudenhancementusingaggregateselectivestatespace modeling. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW) , 180–190URL:https:// api.semanticscholar.org/CorpusID:275932593

    work page 2025

  9. [10]

    Analysis of object detection accu- racy based on the density of 3d point clouds for deep learning- based shipyard datasets

    Jung, K.S., Lee, D.K., 2025. Analysis of object detection accu- racy based on the density of 3d point clouds for deep learning- based shipyard datasets. International Journal of Naval Architec- tureandOceanEngineeringURL:https://api.semanticscholar.org/ CorpusID:276457968

    work page 2025

  10. [11]

    Pointpillars: Fast encoders for object detection from point clouds

    Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O., 2018. Pointpillars: Fast encoders for object detection from point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 12689–12697URL:https://api. semanticscholar.org/CorpusID:55701967

    work page 2018

  11. [12]

    Li, M., Yuan, J., Chen, S., Zhang, L., Zhu, A., Chen, X., Chen, T.,

    work page

  12. [13]

    (Eds.), Advances in Neural InformationProcessingSystems,CurranAssociates,Inc..pp.47242– 47260

    3det-mamba: Causal sequence modelling for end-to-end 3d object detection, in: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C. (Eds.), Advances in Neural InformationProcessingSystems,CurranAssociates,Inc..pp.47242– 47260. URL:https://proceedings.neurips.cc/paper_files/paper/ 2024/file/547108084f0c2af39b956f8eadb75d1b-Pa...

    work page doi:10.52202/079017-1497 2024

  13. [14]

    Pdm-ssd: Single-stage 3d object detector with point dilation mecha- nism

    Liang, A., Hua, H., Fang, J., Chen, W., Zhao, H., Wang, G., 2024a. Pdm-ssd: Single-stage 3d object detector with point dilation mecha- nism. ArXiv abs/2406.00714. doi:10.2139/ssrn.4687799

    work page doi:10.2139/ssrn.4687799

  14. [15]

    Boosting 3d point-based object detection by reducing information loss caused by discontinuous receptive fields

    Liang, A., Hua, H., Fang, J., Zhao, H., Liu, T., 2024b. Boosting 3d point-based object detection by reducing information loss caused by discontinuous receptive fields. Int. J. Appl. Earth Obs. Geoinforma- tion 132, 104049. URL:https://api.semanticscholar.org/CorpusID: 271730018

    work page

  15. [16]

    Pointmamba:Asimplestatespacemodelforpointcloudanal- ysis

    Liang,D.,Zhou,X.,Xu,W.,Zhu,X.,Zou,Z.,Ye,X.,Tan,X.,Bai,X., 2024c. Pointmamba:Asimplestatespacemodelforpointcloudanal- ysis. ArXiv abs/2402.10739. URL:https://api.semanticscholar. org/CorpusID:267740688

    work page arXiv

  16. [17]

    Hypermamba: A spectral-spatial adaptive mamba for hyperspectral image classifi- cation

    Liu, Q., Yue, J., Fang, Y., Xia, S., Fang, L., 2024a. Hypermamba: A spectral-spatial adaptive mamba for hyperspectral image classifi- cation. IEEE Transactions on Geoscience and Remote Sensing 62, 1–14. doi:10.1109/TGRS.2024.3482473

    work page doi:10.1109/tgrs.2024.3482473 2024

  17. [18]

    A unified voxel diffusion module for point cloud 3d object detection

    Liu, Q., Zhao, D., Dong, Y., Shang, L., Xiao, L., Wang, J., Zhao, K., Lu, D., Zhu, Q., 2025a. A unified voxel diffusion module for point cloud 3d object detection. ArXiv abs/2508.16069. URL: https://api.semanticscholar.org/CorpusID:280708543

    work page arXiv

  18. [19]

    Hyperspectral image super-resolution based on mamba and bidirec- tionalfeaturefusionnetwork

    Liu, T., Pu, X., Shi, Y., Liu, Y., Chen, G., Sui, X., Chen, Q., 2025b. Hyperspectral image super-resolution based on mamba and bidirec- tionalfeaturefusionnetwork. ExpertSystemswithApplications285, 127905. URL:https://www.sciencedirect.com/science/article/pii/ S0957417425015271, doi:https://doi.org/10.1016/j.eswa.2025.127905

    work page doi:10.1016/j.eswa.2025.127905 2025

  19. [20]

    Lion:Lineargrouprnnfor3dobjectdetectioninpointclouds

    Liu,Z.,Hou,J.,Wang,X.,Ye,X.,Wang,J.,Zhao,H.,Bai,X.,2024b. Lion:Lineargrouprnnfor3dobjectdetectioninpointclouds. ArXiv abs/2407.18232. URL:https://api.semanticscholar.org/CorpusID: 271431965

    work page arXiv

  20. [21]

    Group-free 3d object detection via transformers

    Liu, Z., Zhang, Z., Cao, Y., Hu, H., Tong, X., 2021. Group-free 3d object detection via transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , 2929–2938URL:https: //api.semanticscholar.org/CorpusID:232478413

    work page 2021

  21. [22]

    Exploring token serialization for mamba-based lidar point cloud segmentation

    Lu, D., Gao, K., Li, J., Zhang, D., Xu, L., 2025. Exploring token serialization for mamba-based lidar point cloud segmentation. IEEE Transactions on Geoscience and Remote Sensing 63, 1–14. URL: https://api.semanticscholar.org/CorpusID:281127906

    work page 2025

  22. [23]

    Mao,J.,Xue,Y.,Niu,M.,Bai,H.,Feng,J.,Liang,X.,Xu,H.,Xu,C.,

    work page

  23. [24]

    2021 IEEE/CVF International Conference on Computer Vision (ICCV) , 3144– 3153URL:https://api.semanticscholar.org/CorpusID:237421262

    Voxel transformer for 3d object detection. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , 3144– 3153URL:https://api.semanticscholar.org/CorpusID:237421262

    work page 2021

  24. [25]

    Mambafusion: State- space model-driven object-scene fusion for multi-modal 3d object detection

    Ning, T., Lu, K., Jiang, X., Xue, J., 2025. Mambafusion: State- space model-driven object-scene fusion for multi-modal 3d object detection. Pattern Recognition URL:https://api.semanticscholar. org/CorpusID:283485417

    work page 2025

  25. [26]

    Ensemble deep learning-enabled single-shot composite structured illumination mi- croscopy (edl-csim)

    Qian, J., Wang, C., Wu, H., Chen, Q., Zuo, C., 2024. Ensemble deep learning-enabled single-shot composite structured illumination mi- croscopy (edl-csim). PhotoniX 1 1. doi:10.1186/s43074-024-00139-2

    work page doi:10.1186/s43074-024-00139-2 2024

  26. [27]

    Center- point transformer for bev object detection with automotive radar

    Saini, L., Su, Y., Tercan, H., Meisen, T., 2024. Center- point transformer for bev object detection with automotive radar. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , 4451–4460URL:https://api. semanticscholar.org/CorpusID:272915648

    work page 2024

  27. [28]

    Pv-rcnn: Point-voxel feature set abstraction for 3d object detection

    Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H., 2019. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. 2020IEEE/CVFConferenceonComputerVisionandPatternRecog- nition (CVPR) , 10526–10535URL:https://api.semanticscholar. org/CorpusID:209516193

    work page 2019

  28. [29]

    Pointrcnn:3dobjectproposalgenera- tion and detection from point cloud

    Shi,S.,Wang,X.,Li,H.,2018. Pointrcnn:3dobjectproposalgenera- tion and detection from point cloud. 2019 IEEE/CVF Conference on Page 10 of 11 Computer Vision and Pattern Recognition (CVPR) , 770–779URL: https://api.semanticscholar.org/CorpusID:54607410

    work page 2018

  29. [30]

    Openpcdet: An open-source toolbox for 3d object detection from point clouds.https://github.com/open-mmlab/ OpenPCDet

    Team, O.D., 2020. Openpcdet: An open-source toolbox for 3d object detection from point clouds.https://github.com/open-mmlab/ OpenPCDet

    work page 2020

  30. [31]

    Mam- baocc: Visual state space model for bev-based occupancy prediction with local adaptive reordering

    Tian, Y., Bai, S., Luo, Z., Wang, Y., Lv, Y., Wang, F.Y., 2024. Mam- baocc: Visual state space model for bev-based occupancy prediction with local adaptive reordering. ArXiv abs/2408.11464. URL:https: //api.semanticscholar.org/CorpusID:271916087

    work page arXiv 2024

  31. [32]

    Pointpainting: Sequential fusion for 3d object detection

    Vora, S., Lang, A.H., Helou, B., Beijbom, O., 2019. Pointpainting: Sequential fusion for 3d object detection. 2020 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , 4603– 4611URL:https://api.semanticscholar.org/CorpusID:208248084

    work page 2019

  32. [33]

    Bridging cross- modal sparsity via adaptive pillar propagation and hierarchical multi-granularity feature distillation for 3d object detection

    Wan, R., Meng, W., Zhao, T., Lu, W., 2026. Bridging cross- modal sparsity via adaptive pillar propagation and hierarchical multi-granularity feature distillation for 3d object detection. Neu- rocomputing 665, 132195. URL:https://www.sciencedirect. com/science/article/pii/S092523122502867X,doi:https://doi.org/10. 1016/j.neucom.2025.132195

    work page arXiv 2026

  33. [34]

    Dsvt: Dynamic sparse voxel transformer with rotated sets

    Wang, H., Shi, C., Shi, S., Lei, M., Wang, S., He, D., Schiele, B., Wang, L., 2023a. Dsvt: Dynamic sparse voxel transformer with rotated sets. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 13520–13529URL:https://api. semanticscholar.org/CorpusID:255942736

    work page 2023

  34. [35]

    Local-to- global structure prior guided high-precision point cloud registration framework based on fpp

    Wang, X., Chen, X., Han, J., Zhang, Y., Zheng, D., 2023b. Local-to- global structure prior guided high-precision point cloud registration framework based on fpp. Measurement 214, 112840. URL:https: //www.sciencedirect.com/science/article/pii/S0263224123004049, doi:https://doi.org/10.1016/j.measurement.2023.112840

    work page doi:10.1016/j.measurement.2023.112840 2023

  35. [36]

    Hybrid transformer- mambamodelfor3dsemanticsegmentation

    Wang, X., Hou, J., Liu, Z., Zhu, Y., 2025. Hybrid transformer- mambamodelfor3dsemanticsegmentation. 2025IEEE/RSJInterna- tional Conference on Intelligent Robots and Systems (IROS) , 2217– 2223URL:https://api.semanticscholar.org/CorpusID:280018036

    work page 2025

  36. [37]

    Pointramba: A hybrid transformer-mamba framework for point cloud analysis

    Wang, Z., Chen, Z., Wu, Y., Zhao, Z., Zhou, L., Xu, D., 2024. Pointramba: A hybrid transformer-mamba framework for point cloud analysis. ArXiv abs/2405.15463. URL:https://api. semanticscholar.org/CorpusID:270045264

    work page arXiv 2024

  37. [38]

    A survey of deep learning based radar and vision fusion for 3d object detection in autonomous driving

    Wu, D., Yang, F., Xu, B., Liao, P., Liu, B., 2024. A survey of deep learning based radar and vision fusion for 3d object detection in autonomous driving. ArXiv abs/2406.00714. URL:https://api. semanticscholar.org/CorpusID:270217797

    work page arXiv 2024

  38. [39]

    Mine-ssd: Dual-threshold set abstraction and radius-adaptive group- ing for 3d object detection in open-pit mines

    Xie, Z., Zhao, Y., Zhang, F., Luo, B., Hu, W., Qiu, T., 2025. Mine-ssd: Dual-threshold set abstraction and radius-adaptive group- ing for 3d object detection in open-pit mines. Neurocomputing 657,131507. URL:https://www.sciencedirect.com/science/article/ pii/S0925231225021794, doi:https://doi.org/10.1016/j.neucom.2025. 131507

    work page doi:10.1016/j.neucom.2025 2025

  39. [40]

    Second: Sparsely embedded convo- lutional detection

    Yan, Y., Mao, Y., Li, B., 2018. Second: Sparsely embedded convo- lutional detection. Sensors (Basel, Switzerland) 18. URL:https: //api.semanticscholar.org/CorpusID:52957856

    work page 2018

  40. [41]

    Chang, Leonidas J

    Yang, Z., Sun, Y., Liu, S., Jia, J., 2020. 3dssd: Point-based 3d single stage object detector, in: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11037–11045. doi:10. 1109/CVPR42600.2020.01105

    work page arXiv 2020

  41. [42]

    Voxel mamba: Group-free state space models for point cloud based 3d object detection

    Zhang, G., Fan, L., He, C., Lei, Z., Zhang, Z., Zhang, L., 2024. Voxel mamba: Group-free state space models for point cloud based 3d object detection. ArXiv abs/2406.10700. URL:https://api. semanticscholar.org/CorpusID:270559452

    work page arXiv 2024

  42. [43]

    Sfnet: A dual-enhanced rgbt tracker via global-local modality refine- ment and frequency-spatial cross-modal fusion

    Zhang, G., Ji, X., Qiu, B., Cai, Y., Liu, Y., Sui, X., Chen, Q., 2025a. Sfnet: A dual-enhanced rgbt tracker via global-local modality refine- ment and frequency-spatial cross-modal fusion. Optics and Lasers in Engineering 194, 109201. URL:https://www.sciencedirect. com/science/article/pii/S0143816625003860,doi:https://doi.org/10. 1016/j.optlaseng.2025.109201

    work page arXiv 2025

  43. [44]

    As- cformer: An adaptive strucure-aware cascaded transformer for 3d object detection

    Zhang, X., Li, X., Zhou, M., Gan, M., Chen, C.L.P., 2025b. As- cformer: An adaptive strucure-aware cascaded transformer for 3d object detection. IEEE Transactions on Circuits and Systems for VideoTechnologyURL:https://api.semanticscholar.org/CorpusID: 281498966

    work page

  44. [45]

    Enhanc- ing point cloud feature utilization for 3d object detection

    Zhou, G., Jiang, F., Dong, R., Wang, M., Xu, P., 2025. Enhanc- ing point cloud feature utilization for 3d object detection. IEEE Signal Processing Letters 32, 4254–4258. URL:https://api. semanticscholar.org/CorpusID:282634331. Page 11 of 11

    work page 2025