HybridTM: Combining Transformer and Mamba for 3D Semantic Segmentation

Jinghua Hou; Xinyu Wang; Yingying Zhu; Zhe Liu

REVIEW 3 major objections 6 minor 49 references

HybridTM: Combining Transformer and Mamba for 3D Semantic Segmentation

T0 review · 3 major / 6 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 HybridTM's Inner Layer Hybrid Strategy fuses Transformer attention and Mamba in each layer, and the paper reports state-of-the-art mIoU on ScanNet, ScanNet200, nuScenes, and S3DIS.

desk verdict A competent incremental hybrid architecture with plausible SOTA scores, but the ablations don't yet prove the gains come from the inner-layer design rather than extra parameters. read the letter →

arxiv 2507.18575 v1 pith:QXKWP52E submitted 2025-07-24 cs.CV

Xinyu Wang , Jinghua Hou , Zhe Liu , Yingying Zhu This is my paper

classification cs.CV

keywords 3DsemanticsegmentationpointcloudTransformerMambastatespacemodelhybridarchitectureInnerLayerStrategyScanNet

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 resolve a trade-off in 3D semantic segmentation: Transformer attention captures precise local relationships but costs quadratic computation, while Mamba processes long sequences in linear time but produces weaker 3D features. The proposed resolution, HybridTM, is a UNet-style network whose hybrid layers first apply multi-head self-attention to small groups of serialized voxels and then apply bidirectional Mamba to larger groups, followed by a feed-forward network. The paper reports that this Inner Layer Hybrid Strategy outperforms attention-only, Mamba-only, and outer hybrid arrangements, and that HybridTM reaches 77.8 mIoU on ScanNet, 36.5 on ScanNet200, 80.9 on nuScenes, and 72.1 on S3DIS. If these numbers hold, the method offers a practical way to get both local detail and long-range context in one efficient layer.

What carries the argument

The load-bearing component is the Inner Layer Hybrid Strategy, defined inside a single Hybrid Layer. Input voxels are serialized by space-filling curves; the layer splits them into non-overlapping small groups of size 1024 for multi-head self-attention, restores the sequence, splits it into larger groups of size 4096 for bidirectional Mamba, restores it again, and closes with an FFN, all with residual connections and xCPE positional encoding. This one-layer ordering is what lets attention supply high-quality local features before Mamba performs efficient global modeling.

What would settle it

Run HybridTM against an attention-only and a Mamba-only variant with matched parameter counts and FLOPs (for example, by widening channels or adding layers to the single-operator baselines). If the matched versions match the 77.8 mIoU result, the hybrid mechanism is not the source of the gain; if a gap persists, the paper's explanation is supported.

Watch

Extended reading notes

Core claim

HybridTM is presented as the first hybrid Transformer-Mamba architecture for 3D semantic segmentation. The central technical claim is that attention and Mamba should be fused inside each layer rather than stacked as two types of layers: attention first enriches small local groups with high-quality fine-grained features, then Mamba consumes those features on groups four times larger to model long-range dependencies at linear cost, with an FFN fusing the two. The paper argues that Mamba's 3D feature degradation comes from weak relative-spatial modeling in serialized voxel sequences, and that feeding it attention-refined local features fixes this. Supporting evidence is the ablation sequence on ScanNet: attention-only scores 77.1 mIoU, Mamba-only 76.9, the combined inner-layer design 77.8; placing Mamba before attention drops to 77.5, and outer-layer stacking drops to 77.1-77.4. Against published methods, HybridTM reports the best mIoU on ScanNet (77.8), ScanNet200 (36.5), and nuScenes (80.9).

Load-bearing premise

The load-bearing assumption is that the reported gains over attention-only and Mamba-only baselines come from the inner-layer hybrid design, not from the extra parameters or compute the combined layer adds.

Editorial extensions

If this is right

Attention-before-Mamba ordering is a concrete design rule: reversing it costs 0.3 mIoU and outer stacking costs 0.4-0.7 mIoU relative to the proposed inner-layer order.
Mamba-based 3D models can recover local feature quality by receiving attention-processed features, so hybrid layers are a viable alternative to pure Mamba backbones.
The same hybrid layer recipe transfers from indoor dense scenes (ScanNet) to outdoor large-scale LiDAR scenes (nuScenes), suggesting the local/global division is not dataset-specific.
Long-range modeling is achieved at linear cost because Mamba sees groups of 4096 voxels, four times the attention group size, while the attention step keeps quadratic cost confined to small groups.

Reading between the lines

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

If the ordering principle is general, it should extend to other sparse serialized inputs such as streaming LiDAR sweeps or medical volumes, but the paper does not test those settings.
The ablation tables do not match parameter count or compute across the three configurations, so part of the 0.7-0.9 point gain may be attributable to the combined layer's extra capacity rather than to the hybrid mechanism itself; a capacity-matched rerun would settle this.
The choice of group sizes (1024 for attention, 4096 for Mamba) is not swept, so the exact trade-off between local detail and global context is likely tunable and may yield further gains on larger datasets.
Because the encoder produces multiscale hybrid features through a standard UNet decoder, the backbone could be lifted to other point-cloud tasks such as object detection or instance segmentation, an untested extrapolation.

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.

Desk Editor's Note

read the letter

HybridTM is a clean, competent engineering contribution: it inserts a Mamba block after a grouped attention block inside each layer of a Point Transformer v3-style network and reports small SOTA gains on ScanNet, ScanNet200, and nuScenes. The thing to know before reading is that the central causal claim—that the inner-layer hybrid design, not extra parameters, drives the gains—is not yet supported by the experiments.

What's genuinely new is the specific combination: attention and Mamba inside the same layer at a fine granularity for 3D segmentation. The design is plausible and clearly explained. The paper does several things right: it compares outer vs inner hybrid strategies, tests operator order, runs standard benchmarks, and reports consistent behavior across indoor and outdoor data. The writing is straightforward, and the authors are honest that Mamba alone underperforms. The xCPE and training recipe are borrowed from PTv3, which is fine, but it means the contribution is incremental.

The soft spots are real and need airing. First, the ablations in Table V are capacity-confounded. The full hybrid layer contains both an attention block and a Mamba block plus an FFN; the single-operator variants drop one of those blocks. The 0.7–0.9 mIoU gains could just be extra parameters. Table VI has the same issue and reports no parameter counts or FLOPs. Second, every number is from a single run. ScanNet validation mIoU for such models varies by roughly 0.3–0.5 across seeds, so the 0.3 margin over PTv3 is within noise, and even the 0.7 ablation margin is not safely outside it. Third, the 'first hybrid architecture for 3D semantic segmentation' claim is asserted rather than demonstrated; the related work only cites a few NLP/2D hybrids and does not engage with any prior 3D hybrid attempts that may exist. Fourth, the S3DIS result (72.1) is not SOTA—Swin3D gets 72.5—so the abstract's phrasing slightly overreaches. These are all addressable.

This is not a bad paper. For a reader in 3D segmentation, it is a reasonable entry point to hybrid designs, and the empirical scores, if they hold across seeds, are a useful data point. But the paper needs capacity-matched ablations, multiple seeds, and released code before the attribution claim can be taken seriously. It deserves peer review, and I would send it, with reviewers specifically asked to scrutinize capacity and noise.

Referee Report

3 major / 6 minor

Summary. HybridTM is a UNet-style 3D semantic segmentation architecture in which each encoder/decoder stage uses a 'Hybrid Layer' containing an xCPE, a windowed multi-head self-attention block (group size 1024), a bidirectional Mamba block (group size 4096), and an FFN. The authors call this the Inner Layer Hybrid (IL) strategy, in contrast to the outer (stage-level) hybrid used in 2D vision. The paper reports 77.8 mIoU on ScanNet, 36.5 on ScanNet200, 80.9 on nuScenes, and 72.1 on S3DIS, claiming state-of-the-art performance, and ablates the attention/Mamba components and the hybrid placement strategy.

Significance. The paper addresses a timely question—whether and how attention and Mamba operators can be combined for 3D point cloud segmentation. The proposed IL strategy is a reasonable design, and the multi-dataset evaluation follows a standard protocol (PTv3's training recipe). If the attribution of the gains to the IL design were rigorously established, the work would be a solid incremental contribution. Currently, however, the central causal claim is under-supported because the ablations do not match model capacity or computational cost, and all numbers are single runs without variance, which is especially problematic given that the margins over baselines are small (0.3–0.9 mIoU).

major comments (3)

[IV-D, Table V] The component ablation does not control for model capacity or compute. Rows I and II each remove one operator block (MSA or BiMamba) from every hybrid layer, while row III keeps both blocks plus the FFN. The 0.7–0.9 mIoU gains of row III could therefore come from roughly the combined parameters and FLOPs of the two operators rather than from a genuine synergy. Please add capacity-matched baselines (e.g., single-operator variants with wider channels or more heads to match total parameters) and report Params and FLOPs for each configuration.
[IV-C, IV-D (Tables I–VI)] All results are reported as single-point numbers with no variance information. On ScanNet validation, typical run-to-run variation for these models is on the order of 0.3–0.5 mIoU, which is comparable to the 0.3 margin over PTv3 (Table I) and the 0.3–0.5 differences in Table VI. Report means and standard deviations over at least three seeds for the main comparisons and the key ablations, and state how the reported numbers were selected.
[IV-D, Table VI] The hybrid strategy comparison is underspecified. It is not clear how many layers are Mamba vs. attention in the Outer Strategy rows, whether the total depth, group sizes, and training epochs are identical across all four rows, or whether the only change is the order/placement. Without this information and without variance estimates, the claim that IL outperforms Outer by 0.4–0.5 mIoU is not supported. Please provide complete configuration details and statistical evidence.

minor comments (6)

[III-A (Eqs. (1)–(2))] There are typos: 'Parition' should be 'Partition' in both equations, and the intermediate F' is reused for both the output of the attention layer and the input of the Mamba layer; consider distinct variable names.
[Tables I–IV] Several table typos: 'Pont Mamba' should be 'Point Mamba' (Table I), 'Cylender3D' should be 'Cylinder3D' (Table III), 'Aera5' should be 'Area5' (Table IV), and 'Scannet' should be 'ScanNet' throughout.
[Table IV] The row 'Serialized Point Mamba [14]' cites the wrong reference; Serialized Point Mamba is reference [26], not [14].
[Tables I–III] The column header 'Present at' is confusing; it should be 'Venue' or 'Publication'.
[Abstract and IV-C] No timing or FLOPs are reported, even though the abstract motivates the design by linear-complexity Mamba processing; adding a computational cost comparison would help evaluate the efficiency claim.
[Abstract and III] The claim of being 'the first hybrid architecture that integrates Transformer and Mamba for 3D semantic segmentation' is not justified by the literature review; several concurrent or prior works on hybrid Mamba-transformer models for point clouds may exist. Please either provide a more comprehensive search of related work or soften this claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical benchmark results measured on external datasets, not derived from the method's definitions; the ablation capacity-matching gap is a validity concern, not a circularity.

full rationale

The paper makes no analytic derivation that could reduce to its inputs. Its central claims are measured mIoU scores on fixed, external benchmarks (ScanNet, ScanNet200, nuScenes, S3DIS) reported in Tables I-IV, and these numbers are not fitted to or defined in terms of the claimed conclusions. The ablations in Tables V and VI compare architectural configurations under stated training settings; even though they do not control parameter count or compute, that is a possible confound in the empirical attribution to the Inner Layer Hybrid design, not a circular step, because the ablations are measurements rather than consequences of the definitions. The xCPE module, loss function, and augmentation strategy are explicitly adopted from Point Transformer V3 by citation to external work and are not equivalent to the paper's target claim. References [4] and [40] include the present authors, but they are cited only as related work on Mamba and 3D object detection and do not carry the load of justifying HybridTM's design or results. The 'first hybrid architecture' claim is a historical priority assertion, not a derived quantity. No self-definition, fitted-input-called-prediction, or self-citation chain that forces the reported outcomes was found.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a set of architectural choices and domain assumptions rather than mathematical derivation. The free parameters are hand-selected and the assumptions are tested only indirectly through ablations on a single dataset.

free parameters (6)

Attention group size L = 1024
Chosen uniformly across all stages for local attention; no sensitivity analysis is provided.
Mamba group size K = 4096
Chosen to represent 'large groups' for global modeling; no ablation on this value.
Encoder layer counts = [2,2,2,6,2]
Standard UNet-style configuration; not ablated.
Decoder layer counts = [2,2,2,2]
Mirrors encoder; not ablated.
Batch size = 12
Chosen for 4 RTX 3090 GPUs; not tuned.
Epochs per dataset = 800/3000/50
Follows common practice; no justification given.

assumptions (3)

domain assumption Space-filling curve serialization preserves spatial locality for grouped attention and Mamba.
The method relies on serialization order from Point Transformer V3; no analysis of order sensitivity is provided.
domain assumption Attention on small groups captures fine-grained local features, while Mamba on large groups captures long-range dependencies.
This is the central design assumption, only indirectly tested via the ablation in Table V.
domain assumption Mamba has weaker feature representation than attention for 3D features.
Motivates the hybrid design; supported by a 0.2 point gap in the authors' own ablation, not by an independent benchmark.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HybridTM: Combining Transformer and Mamba for 3D Semantic Segmentation." pith.science (2026). https://pith.science/paper/QXKWP52E

@misc{pith2026250718575,
  author       = {Pith},
  title        = {Pith review of: HybridTM: Combining Transformer and Mamba for 3D Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QXKWP52E}},
  note         = {Machine review of arXiv:2507.18575}
}

read the original abstract

Transformer-based methods have demonstrated remarkable capabilities in 3D semantic segmentation through their powerful attention mechanisms, but the quadratic complexity limits their modeling of long-range dependencies in large-scale point clouds. While recent Mamba-based approaches offer efficient processing with linear complexity, they struggle with feature representation when extracting 3D features. However, effectively combining these complementary strengths remains an open challenge in this field. In this paper, we propose HybridTM, the first hybrid architecture that integrates Transformer and Mamba for 3D semantic segmentation. In addition, we propose the Inner Layer Hybrid Strategy, which combines attention and Mamba at a finer granularity, enabling simultaneous capture of long-range dependencies and fine-grained local features. Extensive experiments demonstrate the effectiveness and generalization of our HybridTM on diverse indoor and outdoor datasets. Furthermore, our HybridTM achieves state-of-the-art performance on ScanNet, ScanNet200, and nuScenes benchmarks. The code will be made available at https://github.com/deepinact/HybridTM.

Figures

Figures reproduced from arXiv: 2507.18575 by the authors.

**Figure 2.** The illustration of HybridTM, which consists of HybridTM encoders, HybridTM decoders, down-sampling, up-sampling, and a classification head. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The illustration of Hybrid Layer. The Hybrid Layer contains an xCPE, an attention layer, a Mamba layer, and an FFN layer. First, we divide the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Comparison of different hybrid strategies. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Comparison of Point Transformer V3 (b) and HybridTM (c) on the ScanNet and nuScenes validation set. (a) is the ground truth. The first and [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 38 canonical work pages

[1]

A comparative study of real-time semantic segmentation for autonomous driving,

M. Siam, M. Gamal, M. Abdel-Razek, S. Yogamani, M. Jagersand, and H. Zhang, “A comparative study of real-time semantic segmentation for autonomous driving,” in Proceedings of the IEEE conference on computer vision and pattern recognition workshops , 2018, pp. 587– 597

work page 2018
[2]

Mask-based panoptic lidar segmentation for autonomous driving,

R. Marcuzzi, L. Nunes, L. Wiesmann, J. Behley, and C. Stachniss, “Mask-based panoptic lidar segmentation for autonomous driving,” IEEE Robotics and Automation Letters , vol. 8, no. 2, pp. 1141–1148, 2023

work page 2023
[3]

Pointmoseg: Sparse tensor-based end-to-end moving-obstacle segmentation in 3-d lidar point clouds for autonomous driving,

Y . Sun, W. Zuo, H. Huang, P. Cai, and M. Liu, “Pointmoseg: Sparse tensor-based end-to-end moving-obstacle segmentation in 3-d lidar point clouds for autonomous driving,” IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 510–517, 2020

work page 2020
[4]

Epnet++: Cascade bi-directional fusion for multi-modal 3d object detection,

Z. Liu, T. Huang, B. Li, X. Chen, X. Wang, and X. Bai, “Epnet++: Cascade bi-directional fusion for multi-modal 3d object detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 7, pp. 8324–8341, 2022

work page 2022
[5]

Indoor semantic segmentation for robot navigat- ing on mobile,

W. Kim and J. Seok, “Indoor semantic segmentation for robot navigat- ing on mobile,” in 2018 Tenth International Conference on Ubiquitous and Future Networks (ICUFN) . IEEE, 2018, pp. 22–25

work page 2018
[6]

Multi-view incremental segmentation of 3-d point clouds for mobile robots,

J. Chen, Y . K. Cho, and Z. Kira, “Multi-view incremental segmentation of 3-d point clouds for mobile robots,” IEEE Robotics and Automation Letters, vol. 4, no. 2, pp. 1240–1246, 2019

work page 2019
[7]

Semantickitti: A dataset for semantic scene under- standing of lidar sequences,

J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stach- niss, and J. Gall, “Semantickitti: A dataset for semantic scene under- standing of lidar sequences,” in Porc. of IEEE Intl. Conf. on Computer Vision, 2019, pp. 9297–9307

work page 2019
[8]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Proc. of Advances in Neural Information Processing Systems , vol. 30, 2017

work page 2017

Show all 49 references

[9]

Point transformer,

H. Zhao, L. Jiang, J. Jia, P. H. Torr, and V . Koltun, “Point transformer,” in Porc. of IEEE Intl. Conf. on Computer Vision , 2021, pp. 16 259– 16 268

2021
[10]

Patchformer: An efficient point transformer with patch attention,

C. Zhang, H. Wan, X. Shen, and Z. Wu, “Patchformer: An efficient point transformer with patch attention,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2022, pp. 11 799–11 808

2022
[11]

Swin3d: A pretrained transformer backbone for 3d indoor scene understanding,

Y .-Q. Yang, Y .-X. Guo, J.-Y . Xiong, Y . Liu, H. Pan, P.-S. Wang, X. Tong, and B. Guo, “Swin3d: A pretrained transformer backbone for 3d indoor scene understanding,” arXiv preprint arXiv:2304.06906 , 2023

2023 arXiv
[12]

Octformer: Octree-based transformers for 3d point clouds,

P.-S. Wang, “Octformer: Octree-based transformers for 3d point clouds,” ACM Transactions ON Graphics , vol. 42, no. 4, pp. 1–11, 2023

2023
[13]

Point transformer v2: Grouped vector attention and partition-based pooling,

X. Wu, Y . Lao, L. Jiang, X. Liu, and H. Zhao, “Point transformer v2: Grouped vector attention and partition-based pooling,” Proc. of Advances in Neural Information Processing Systems , vol. 35, pp. 33 330–33 342, 2022

2022
[14]

Point transformer v3: Simpler faster stronger,

X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y . Qiao, W. Ouyang, T. He, and H. Zhao, “Point transformer v3: Simpler faster stronger,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2024, pp. 4840–4851

2024
[15]

Fast point transformer,

C. Park, Y . Jeong, M. Cho, and J. Park, “Fast point transformer,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2022, pp. 16 949–16 958

2022
[16]

Stratified transformer for 3d point cloud segmentation,

X. Lai, J. Liu, L. Jiang, L. Wang, H. Zhao, S. Liu, X. Qi, and J. Jia, “Stratified transformer for 3d point cloud segmentation,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2022, pp. 8500–8509

2022
[17]

Spherical transformer for lidar-based 3d recognition,

X. Lai, Y . Chen, F. Lu, J. Liu, and J. Jia, “Spherical transformer for lidar-based 3d recognition,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2023, pp. 17 545–17 555

2023
[18]

Mamba: Linear-time sequence modeling with selective state spaces,

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

2023 arXiv
[19]

Vision mamba: Efficient visual representation learning with bidirectional state space model,

L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” in Proc. of Intl. Conf. on Machine Learning , 2024

2024
[20]

Vmamba: Visual state space model,

Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, and Y . Liu, “Vmamba: Visual state space model,” Proc. of Advances in Neural Information Processing Systems , 2024

2024
[21]

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,”arXiv preprint arXiv:2403.09338, 2024

2024 arXiv
[22]

Efficientvmamba: Atrous selective scan for light weight visual mamba,

X. Pei, T. Huang, and C. Xu, “Efficientvmamba: Atrous selective scan for light weight visual mamba,” arXiv preprint arXiv:2403.09977, 2024

2024 arXiv
[23]

Vm-unet: Vision mamba unet for medical image segmentation,

J. Ruan and S. Xiang, “Vm-unet: Vision mamba unet for medical image segmentation,” arXiv preprint arXiv:2402.02491 , 2024

2024 arXiv
[24]

Point mamba: A novel point cloud backbone based on state space model with octree-based ordering strategy,

J. Liu, R. Yu, Y . Wang, Y . Zheng, T. Deng, W. Ye, and H. Wang, “Point mamba: A novel point cloud backbone based on state space model with octree-based ordering strategy,”arXiv preprint arXiv:2403.06467, 2024

2024 arXiv
[25]

Point cloud mamba: Point cloud learning via state space model,

T. Zhang, X. Li, H. Yuan, S. Ji, and S. Yan, “Point cloud mamba: Point cloud learning via state space model,” arXiv preprint arXiv:2403.00762, 2024

2024 arXiv
[26]

Serialized point mamba: A serialized point cloud mamba segmentation model,

T. Wang, W. Wen, J. Zhai, K. Xu, and H. Luo, “Serialized point mamba: A serialized point cloud mamba segmentation model,” arXiv preprint arXiv:2407.12319, 2024

2024 arXiv
[27]

3d semantic segmentation with submanifold sparse convolutional networks,

B. Graham, M. Engelcke, and L. Van Der Maaten, “3d semantic segmentation with submanifold sparse convolutional networks,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2018, pp. 9224–9232

2018
[28]

Scannet: Richly-annotated 3d reconstructions of indoor scenes,

A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor scenes,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition, 2017, pp. 5828–5839

2017
[29]

nuscenes: A multimodal dataset for autonomous driving,

H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2020, pp. 11 621– 11 631

2020
[30]

Jamba: A hybrid transformer-mamba language model,

O. Lieber, B. Lenz, H. Bata, G. Cohen, J. Osin, I. Dalmedi- gos, E. Safahi, S. Meirom, Y . Belinkov, S. Shalev-Shwartz et al. , “Jamba: A hybrid transformer-mamba language model,” arXiv preprint arXiv:2403.19887, 2024

2024 arXiv
[31]

Mambavision: A hybrid mamba- transformer vision backbone,

A. Hatamizadeh and J. Kautz, “Mambavision: A hybrid mamba- transformer vision backbone,” arXiv preprint arXiv:2407.08083, 2024

2024 arXiv
[32]

Map: Unleashing hybrid mamba-transformer vision backbone’s potential with masked autoregressive pretraining,

Y . Liu and L. Yi, “Map: Unleashing hybrid mamba-transformer vision backbone’s potential with masked autoregressive pretraining,” arXiv preprint arXiv:2410.00871, 2024

2024 arXiv
[33]

Maskmamba: A hybrid mamba-transformer model for masked image generation,

W. Chen, L. Niu, Z. Lu, F. Meng, and J. Zhou, “Maskmamba: A hybrid mamba-transformer model for masked image generation,”arXiv preprint arXiv:2409.19937, 2024

2024 arXiv
[34]

Pct: Point cloud transformer,

M.-H. Guo, J.-X. Cai, Z.-N. Liu, T.-J. Mu, R. R. Martin, and S.-M. Hu, “Pct: Point cloud transformer,” Computational Visual Media , vol. 7, pp. 187–199, 2021

2021
[35]

Gaussian radar transformer for semantic segmentation in noisy radar data,

M. Zeller, J. Behley, M. Heidingsfeld, and C. Stachniss, “Gaussian radar transformer for semantic segmentation in noisy radar data,” IEEE Robotics and Automation Letters , vol. 8, no. 1, pp. 344–351, 2022

2022
[36]

Transformers are ssms: Generalized models and efficient algorithms through structured state space duality,

T. Dao and A. Gu, “Transformers are ssms: Generalized models and efficient algorithms through structured state space duality,” in Proc. of Intl. Conf. on Machine Learning , 2024

2024
[37]

Hydra: Bidirectional state space models through generalized matrix mixers,

S. Hwang, A. Lahoti, T. Dao, and A. Gu, “Hydra: Bidirectional state space models through generalized matrix mixers,” arXiv preprint arXiv:2407.09941, 2024

2024 arXiv
[38]

Mim-istd: Mamba-in-mamba for efficient infrared small target detection,

T. Chen, Z. Ye, Z. Tan, T. Gong, Y . Wu, Q. Chu, B. Liu, N. Yu, and J. Ye, “Mim-istd: Mamba-in-mamba for efficient infrared small target detection,” IEEE Transactions on Geoscience and Remote Sensing , 2024

2024
[39]

Omega: Efficient occlusion-aware navigation for air-ground robot in dynamic environments via state space model,

J. Wang, D. Huang, X. Guan, Z. Sun, T. Shen, F. Liu, and H. Cui, “Omega: Efficient occlusion-aware navigation for air-ground robot in dynamic environments via state space model,” IEEE Robotics and Automation Letters, 2024

2024
[40]

Lion: Linear group rnn for 3d object detection in point clouds,

Z. Liu, J. Hou, X. Wang, X. Ye, J. Wang, H. Zhao, and X. Bai, “Lion: Linear group rnn for 3d object detection in point clouds,” Proc. of Advances in Neural Information Processing Systems , 2024

2024
[41]

3d semantic parsing of large-scale indoor spaces,

I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis, M. Fischer, and S. Savarese, “3d semantic parsing of large-scale indoor spaces,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2016, pp. 1534–1543

2016
[42]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in Proc. of Intl. Conf. on Learning Representations , 2017. [Online]. Available: https://api.semanticscholar.org/CorpusID:53592270

2017
[43]

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,” Proc. of Advances in Neural Information Processing Systems , vol. 30, 2017

2017
[44]

4d spatio-temporal convnets: Minkowski convolutional neural networks,

C. Choy, J. Gwak, and S. Savarese, “4d spatio-temporal convnets: Minkowski convolutional neural networks,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2019, pp. 3075– 3084

2019
[45]

O-cnn: Octree-based convolutional neural networks for 3d shape analysis,

P.-S. Wang, Y . Liu, Y .-X. Guo, C.-Y . Sun, and X. Tong, “O-cnn: Octree-based convolutional neural networks for 3d shape analysis,” ACM Transactions ON Graphics , vol. 36, no. 4, pp. 1–11, 2017

2017
[46]

Search- ing efficient 3d architectures with sparse point-voxel convolution,

H. Tang, Z. Liu, S. Zhao, Y . Lin, J. Lin, H. Wang, and S. Han, “Search- ing efficient 3d architectures with sparse point-voxel convolution,” in Proc. of European Conference on Computer Vision . Springer, 2020, pp. 685–702

2020
[47]

Cylindrical and asymmetrical 3d convolution networks for lidar segmentation,

X. Zhu, H. Zhou, T. Wang, F. Hong, Y . Ma, W. Li, H. Li, and D. Lin, “Cylindrical and asymmetrical 3d convolution networks for lidar segmentation,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2021, pp. 9939–9948

2021
[48]

2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network,

R. Cheng, R. Razani, E. Taghavi, E. Li, and B. Liu, “2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network,” in Proc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition , 2021, pp. 12 547–12 556

2021
[49]

Pointnext: Revisiting pointnet++ with improved training and scaling strategies,

G. Qian, Y . Li, H. Peng, J. Mai, H. Hammoud, M. Elhoseiny, and B. Ghanem, “Pointnext: Revisiting pointnet++ with improved training and scaling strategies,” Proc. of Advances in Neural Information Processing Systems, vol. 35, pp. 23 192–23 204, 2022

2022

Pith tools

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

[1] [1]

A comparative study of real-time semantic segmentation for autonomous driving,

M. Siam, M. Gamal, M. Abdel-Razek, S. Yogamani, M. Jagersand, and H. Zhang, “A comparative study of real-time semantic segmentation for autonomous driving,” in Proceedings of the IEEE conference on computer vision and pattern recognition workshops , 2018, pp. 587– 597

work page 2018

[2] [2]

Mask-based panoptic lidar segmentation for autonomous driving,

R. Marcuzzi, L. Nunes, L. Wiesmann, J. Behley, and C. Stachniss, “Mask-based panoptic lidar segmentation for autonomous driving,” IEEE Robotics and Automation Letters , vol. 8, no. 2, pp. 1141–1148, 2023

work page 2023

[3] [3]

Pointmoseg: Sparse tensor-based end-to-end moving-obstacle segmentation in 3-d lidar point clouds for autonomous driving,

Y . Sun, W. Zuo, H. Huang, P. Cai, and M. Liu, “Pointmoseg: Sparse tensor-based end-to-end moving-obstacle segmentation in 3-d lidar point clouds for autonomous driving,” IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 510–517, 2020

work page 2020

[4] [4]

Epnet++: Cascade bi-directional fusion for multi-modal 3d object detection,

Z. Liu, T. Huang, B. Li, X. Chen, X. Wang, and X. Bai, “Epnet++: Cascade bi-directional fusion for multi-modal 3d object detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 7, pp. 8324–8341, 2022

work page 2022

[5] [5]

Indoor semantic segmentation for robot navigat- ing on mobile,

W. Kim and J. Seok, “Indoor semantic segmentation for robot navigat- ing on mobile,” in 2018 Tenth International Conference on Ubiquitous and Future Networks (ICUFN) . IEEE, 2018, pp. 22–25

work page 2018

[6] [6]

Multi-view incremental segmentation of 3-d point clouds for mobile robots,

J. Chen, Y . K. Cho, and Z. Kira, “Multi-view incremental segmentation of 3-d point clouds for mobile robots,” IEEE Robotics and Automation Letters, vol. 4, no. 2, pp. 1240–1246, 2019

work page 2019

[7] [7]

Semantickitti: A dataset for semantic scene under- standing of lidar sequences,

J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stach- niss, and J. Gall, “Semantickitti: A dataset for semantic scene under- standing of lidar sequences,” in Porc. of IEEE Intl. Conf. on Computer Vision, 2019, pp. 9297–9307

work page 2019

[8] [8]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Proc. of Advances in Neural Information Processing Systems , vol. 30, 2017

work page 2017