Pith. sign in

REVIEW 4 major objections 4 minor 46 references

HTMNet: A Hybrid Network with Transformer-Mamba Bottleneck Multimodal Fusion for Transparent and Reflective Objects Depth Completion

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read HTMNet claims state-of-the-art depth completion for transparent and reflective objects by fusing Transformer, CNN, and Mamba features at the network bottleneck.

desk verdict A sensible hybrid architecture with a first Mamba application, but the reported gains are marginal and the Mamba-specific benefit is not isolated. read the letter →

arxiv 2505.20904 v2 pith:RHS7ZQHR submitted 2025-05-27 cs.CV

Guanghu Xie , Yonglong Zhang , Zhiduo Jiang , Yang Liu , Zongwu Xie , Baoshi Cao , Hong Liu This is my paper
classification cs.CV
keywords DepthInpaintingTransparentObjectsReflectiveCompletionStateSpaceModelsTransformerMambaMultimodalFusion
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

Transparent and reflective objects break standard depth sensors, leaving holes and false background readings that harm robotic manipulation. This paper proposes HTMNet, a hybrid network whose encoder uses one Transformer branch for RGB-D images and one CNN branch for raw depth, with a bottleneck fusion module that combines self-attention and a Mamba state-space block. The authors claim that this design achieves state-of-the-art accuracy on the TransCG, ClearGrasp, and STD benchmarks, and that it is the first use of the Mamba architecture in transparent-object depth completion. A reader should care because better completed depth directly feeds downstream grasping and pose-estimation systems.

What carries the argument

The Bottleneck Fusion Module (BFM) is the central mechanism: it adds RGB-D and depth features element-wise, then passes them through a multi-head self-attention block, a Mamba block (a gated state-space model with a 1D convolution, SiLU activation, and linear expansion), and an MLP block. A Multi-Scale Fusion Module (MSFM) in the decoder combines channel attention, spatial attention, and multi-scale depthwise convolutions with a channel-shuffle operation to merge encoder features at different resolutions.

What would settle it

Retrain the model on TransCG with the Mamba block inside the bottleneck replaced by a standard Transformer self-attention block of comparable parameter count and FLOPs, keeping everything else fixed; if $\delta_{1.05}$ stays within the 92.3–92.4 range, the state-space component is not the source of the claimed advantage. Additionally, run the full and ablated configurations across several random seeds to check whether the 92.35-to-92.40 gap is real.

Watch

Extended reading notes

Core claim

On the paper's own terms, HTMNet establishes that a Transformer-Mamba bottleneck can fuse multimodal features more effectively than layer-wise fusion or earlier CNN-Transformer baselines. The quantitative evidence is a set of metrics on three public datasets: on TransCG the full model reaches $\delta_{1.05}=92.40$ with RMSE 0.012; on ClearGrasp real-world it reaches $\delta_{1.05}=86.32$ with RMSE 0.020; on STD it reaches $\delta_{1.05}=91.26$ with RMSE 0.019, each surpassing the compared methods. The paper attributes this to the bottleneck fusion of global self-attention with the selective state-space scan, plus the decoder's multi-scale spatial-channel attention fusion.

Load-bearing premise

The load-bearing premise is that the Transformer-Mamba bottleneck fusion module is what causes the reported performance gain; the paper only ablates the module as a whole and never replaces the Mamba block with a standard block of matched capacity, so the specific role of the state-space model is unverified.

Editorial extensions

If this is right

  • If the claimed gains hold, the Mamba architecture becomes a viable component for transparent-object depth completion, not just for long-sequence tasks.
  • The bottleneck-fusion strategy delivers higher throughput (286 images/sec) than layer-wise fusion (230 images/sec) with comparable accuracy, which matters for real-time robotic perception.
  • The reported state-of-the-art results on TransCG, ClearGrasp, and STD give downstream grasping and manipulation systems a new default depth-completion module.
  • The multi-scale fusion decoder design is separable from the Mamba component and could be reused in other depth-completion or inpainting networks.

Reading between the lines

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

  • The authors do not swap the Mamba block for a matched-capacity Transformer or CNN block, so whether the state-space component itself, rather than the added parameters, drives the 0.05-point $\delta_{1.05}$ gain over the no-BFM ablation is untested.
  • The ablation comparison shows layer-wise fusion reaches a slightly higher $\delta_{1.05}$ (92.51) than bottleneck fusion (92.40), so the bottleneck choice is a speed-accuracy trade-off rather than a strict accuracy improvement.
  • Because Mamba has linear complexity in sequence length, a natural extension would be to test HTMNet at higher input resolutions or on transparent-object video depth completion, where self-attention cost grows quadratically; the paper does not do this.
  • The reported SOTA margins over TDCNet are small (identical RMSE on TransCG), so the practical takeaway may be architectural novelty plus modest metric gains rather than a step-change in accuracy.
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 / 4 minor

Summary. The paper proposes HTMNet, a hybrid encoder-bottleneck-decoder network for depth completion of transparent and reflective objects. The encoder uses a Swin Transformer branch for RGB-D features and a ResNet branch for raw depth; the bottleneck fuses the resulting multimodal features with a module combining multi-head self-attention, a Mamba/SSM block, and an MLP; the decoder uses a multi-scale fusion module with channel and spatial attention. The model is trained with an MSE plus normal-based smoothing loss and evaluated on TransCG, ClearGrasp, and STD against several baselines. The main claims are state-of-the-art (SOTA) performance on these datasets and the first application of the Mamba architecture to transparent object depth completion.

Significance. If the results hold, HTMNet offers a concrete, well-described architecture for a practically important task, and the three-dataset evaluation with multiple baselines is a useful empirical contribution. The paper is clearly written, the method is reproducible in principle from the description, and there is no circularity in the evaluation: the model is trained and tested on public datasets with standard losses. However, the claimed SOTA margins are small and inconsistent across metrics and datasets, and the current experiments do not isolate the Mamba component as the source of the gains. As reported, the central claims are not yet fully supported.

major comments (4)
  1. [Section IV.C.1, Table I] The SOTA claim is not consistent across metrics on TransCG: HTMNet ties TDCNet on RMSE and MAE (0.012 and 0.008), is worse on REL (0.018 vs 0.017), and leads only by 0.15 percentage points in δ1.05 (92.40 vs 92.25). Because the text states that HTMNet 'achieves state-of-the-art performance,' these per-metric comparisons need to be acknowledged and the claim qualified, or the margins must be shown to be statistically meaningful.
  2. [Section IV.C.2, Table II] On ClearGrasp, HTMNet is worse than DITR on RMSE (0.020 vs 0.019) and MAE (0.015 vs 0.012), and the DITR numbers are taken from the official paper rather than retrained under the same protocol as the other baselines, as stated in the text. This unbalanced comparison does not support the sentence that HTMNet 'surpasses many existing state-of-the-art approaches' without explicitly noting the protocol difference and the metric-specific losses.
  3. [Section V, Tables V and VI] The ablations do not isolate the Mamba block. Table VI removes the entire bottleneck fusion module (BFM) as a binary switch, and Table V varies only the number of identical BFM modules; there is no experiment replacing the Mamba block with a matched-capacity Transformer or convolutional block. Since the BFM also contains a multi-head self-attention block and an MLP, the reported numbers do not establish that the Mamba/SSM component, rather than added capacity or the attention/MLP branches, is responsible for the observed differences.
  4. [Section V, Tables I-VI] No repeated runs or standard deviations are reported anywhere, and the reported differences are small; for example, δ1.05 moves 92.21→92.40→92.34 as the number of BFM modules goes from 2 to 4 to 6 in Table V, a non-monotonic pattern consistent with seed noise. Without error bars or significance testing, the central claims that HTMNet is SOTA and that the BFM is beneficial are not statistically supported.
minor comments (4)
  1. [Section IV.C.3 and Table III] The text refers to 'Fif.9' and the table caption reads 'STDS Real-world Dataset'; both should be corrected to 'Fig. 9' and 'STD Real-world Dataset'.
  2. [Tables I and II] Several rows in the tables have missing spaces between numbers, e.g., '0.0190.00890.43' in the TODE-Trans row of Table I and similar entries in Table II; please format all table cells consistently.
  3. [Section III.B, Eq. (5)] The dimensions of the two Mamba-branch activations are described in the text as (B, 2L, C) and (B, L, C), but the equation does not make these shapes explicit, and the relationship between the expansion ratio E and the factor 2 is not stated; a short clarification would improve reproducibility.
  4. [Section V.A, Table IV] The abbreviation 'BF' is used in the table without being defined in the caption; please define it as 'bottleneck fusion'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: HTMNet's claims are empirical evaluations on public benchmarks with standard losses; no derivation reduces to its own inputs.

full rationale

No circularity found. The paper's central claims are empirical: HTMNet is trained end-to-end on public datasets (TransCG, ClearGrasp, STD) with a standard MSE plus normal-smoothing loss (Eq. 9). The components (Swin Transformer encoder, ResNet CNN branch, BFM using self-attention and Mamba SSM, MSFM decoder) are architectural choices, not quantities derived from the target metrics. The BFM equations (3)-(6) and MSFM equations (7)-(8) define the forward computation; loss values are computed against ground-truth depth, and reported metrics are measurements on held-out test sets. No parameter is fitted to a subset of the test data and then renamed a prediction. The comparison tables cite prior methods as external baselines, but the load-bearing evidence is the retraining protocol and measured metrics, not a self-citation chain. The 'first application of Mamba' claim is a novelty statement, not a derivation. The Limitations section explicitly acknowledges a trade-off (degraded precision for non-transparent regions), which is a correctness/robustness concern, not circularity. The absence of error bars and the small margins (e.g., δ1.05 92.40 vs 92.25 in Table I) are statistical-evidence concerns that belong to reproducibility/validity, not to circular reasoning; they do not make the derivation equivalent to its inputs.

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

The central contribution is empirical: an architecture trained with standard losses on public datasets. There are no new physical or mathematical entities. The main hand-tuned elements are structural hyperparameters (module count, expansion ratio, loss weight), and the performance claims rest on the trustworthiness of the public benchmarks and the validity of standard SSM discretization.

free parameters (4)
  • Number of bottleneck fusion modules = 4
    Selected by ablation in Table V; values 2, 4, 6 are tested with nearly identical accuracy, so the choice is hand-tuned.
  • Loss balancing coefficient alpha = not reported
    Introduced in Eq. 9 to weight the normal-based smoothness loss; the paper never gives its value, treating it as an implicit hyperparameter.
  • Mamba block expansion ratio E = not reported
    Used in Eq. 5 to expand channel dimension; no value or search is given.
  • Input resolution = 320x240
    All images are resized to this resolution, as stated in Section IV-B; the sensitivity of the result to this choice is not studied.
assumptions (4)
  • standard math The discretized state-space model equations (Eq. 2) are a valid approximation of the continuous SSM.
    This is standard Mamba/S4 discretization, not introduced by the paper.
  • domain assumption The MSE plus normal-smoothing loss (Eq. 9) is an appropriate training objective for depth completion.
    Adopted from earlier works [24] and [25]; the paper does not justify it beyond citing similarity.
  • domain assumption Ground-truth depth maps in TransCG, ClearGrasp, and STD are reliable enough to train and evaluate on.
    All conclusions depend on dataset label quality, which the paper does not analyze.
  • domain assumption The pretrained Swin Transformer and ResNet backbones provide useful feature representations for RGB-D and depth inputs.
    The encoder relies on these pretrained backbones, whose suitability is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HTMNet: A Hybrid Network with Transformer-Mamba Bottleneck Multimodal Fusion for Transparent and Reflective Objects Depth Completion." pith.science (2026). https://pith.science/paper/RHS7ZQHR

@misc{pith2026250520904,
  author       = {Pith},
  title        = {Pith review of: HTMNet: A Hybrid Network with Transformer-Mamba Bottleneck Multimodal Fusion for Transparent and Reflective Objects Depth Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RHS7ZQHR}},
  note         = {Machine review of arXiv:2505.20904}
}
read the original abstract

Transparent and reflective objects pose significant challenges for depth sensors, resulting in incomplete depth information that adversely affects downstream robotic perception and manipulation tasks. To address this issue, we propose HTMNet, a novel hybrid model integrating Transformer, CNN, and Mamba architectures. The encoder is based on a dual-branch CNN-Transformer framework, the bottleneck fusion module adopts a Transformer-Mamba architecture, and the decoder is built upon a multi-scale fusion module. We introduce a novel multimodal fusion module grounded in self-attention mechanisms and state space models, marking the first application of the Mamba architecture in the field of transparent object depth completion and revealing its promising potential. Additionally, we design an innovative multi-scale fusion module for the decoder that combines channel attention, spatial attention, and multi-scale feature extraction techniques to effectively integrate multi-scale features through a down-fusion strategy. Extensive evaluations on multiple public datasets demonstrate that our model achieves state-of-the-art(SOTA) performance, validating the effectiveness of our approach.

Figures

Figures reproduced from arXiv: 2505.20904 by the authors.

Figure 1
Figure 1. a) illustrates two typical types of errors encountered when [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Depth completion plays a crucial role in dexterous grasping [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. HTMNet Architecture.Our method consists of a dual-branch encoder, a bottleneck fusion module, and a decoder. The Transformer [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Bottleneck fusion module(BFM): Composed of Trans [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Multi-scale fusion module(MSFM): Constructed based on [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 7
Figure 7. Figure 7: Depth Completion Visualizations of Different Models on the TransCG Dataset [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Depth Completion Visualizations of Different Models on the ClearGrasp Real-world Dataset [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Depth Completion Visualizations of Different Models on the STDS Real-world Dataset [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Depth Completion Visualizations on the Different Real-world Dataset [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 34 canonical work pages

  1. [1]

    Light field depth estimation for non-lambertian objects via adaptive cross operator,

    Z. Cui, H. Sheng, D. Yang, S. Wang, R. Chen, and W. Ke, “Light field depth estimation for non-lambertian objects via adaptive cross operator,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 2, pp. 1199–1211, 2023

    2023

  2. [2]

    Diffusion-based depth inpainting for transparent and reflective objects,

    T. Sun, D. Hu, Y. Dai, and G. Wang, “Diffusion-based depth inpainting for transparent and reflective objects,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    work page 2024

  3. [3]

    Transdiff: Diffusion-based method for manipulating transparent objects using a single rgb-d image,

    H. Wang, K. Zhou, B. Gu, Z. Feng, W. Wang, P. Sun, Y. Xiao, J. Zhang, and H. Dong, “Transdiff: Diffusion-based method for manipulating transparent objects using a single rgb-d image,” arXiv preprint arXiv:2503.12779, 2025

    arXiv 2025

  4. [4]

    GAA-TSO: Geometry-Aware Assisted Depth Completion for Transparent and Specular Objects

    Y. Liu, T. Jia, D. Cai, H. Wang, and D. Chen, “Gaa-tso: Geometry-aware assisted depth completion for transparent and specular objects,”arXiv preprint arXiv:2503.17106, 2025

    work page Pith review arXiv 2025

  5. [5]

    Domain randomization-enhanced depth simulation and restoration for perceiving and grasping specular and trans- parent objects,

    Q. Dai, J. Zhang, Q. Li, T. Wu, H. Dong, Z. Liu, P. Tan, and H. Wang, “Domain randomization-enhanced depth simulation and restoration for perceiving and grasping specular and trans- parent objects,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 374–391

    work page 2022

  6. [6]

    Distillgrasp: Integrating features correlation with knowledge distillation for depth completion of transparent objects,

    Y. Huang, J. Chen, N. Michiels, M. Asim, L. Claesen, and W. Liu, “Distillgrasp: Integrating features correlation with knowledge distillation for depth completion of transparent objects,”IEEE Robotics and Automation Letters, 2024

    work page 2024

  7. [7]

    Learning depth estimation for transparent and mirror surfaces,

    A. Costanzino, P. Z. Ramirez, M. Poggi, F. Tosi, S. Mattoccia, and L. Di Stefano, “Learning depth estimation for transparent and mirror surfaces,” inProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, 2023, pp. 9244– 9255

    work page 2023

  8. [8]

    Robust plane detection using depth information from a consumer depth camera,

    Z. Jin, T. Tillo, W. Zou, Y. Zhao, and X. Li, “Robust plane detection using depth information from a consumer depth camera,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 29, no. 2, pp. 447–460, 2017

    work page 2017

Show all 46 references
  1. [9]

    Transparent object depth completion,

    Y. Zhou, W. Peng, Z. Yang, H. Liu, and Y. Sun, “Transparent object depth completion,”arXiv preprint arXiv:2405.15299, 2024

    2024 arXiv

  2. [10]

    Transparent object depth perception network for robotic manipulation based on orientation-aware guidance and texture enhancement,

    Y. Yan, H. Tian, K. Song, Y. Li, Y. Man, and L. Tong, “Transparent object depth perception network for robotic manipulation based on orientation-aware guidance and texture enhancement,”IEEE Transactions on Instrumentation and Measurement, 2024

    2024

  3. [11]

    Segment, lift and fit: Automatic 3d shape labeling from 2d prompts,

    J. Li, T. Sun, Z. Wang, E. Xie, B. Feng, H. Zhang, Z. Yuan, K. Xu, J. Liu, and P. Luo, “Segment, lift and fit: Automatic 3d shape labeling from 2d prompts,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 407–423

    2024

  4. [12]

    Gamba: Marry gaussian splatting with mamba for single view 3d reconstruction,

    Q. Shen, Z. Wu, X. Yi, P. Zhou, H. Zhang, S. Yan, and X. Wang, “Gamba: Marry gaussian splatting with mamba for single view 3d reconstruction,”arXiv preprint arXiv:2403.18795, 2024

    2024 arXiv

  5. [13]

    Decomposed guided dynamic filters for efficient rgb-guided depth completion,

    Y. Wang, Y. Mao, Q. Liu, and Y. Dai, “Decomposed guided dynamic filters for efficient rgb-guided depth completion,” IEEE Transactions on Circuits and Systems for Video Tech- nology, vol. 34, no. 2, pp. 1186–1198, 2023

    2023

  6. [14]

    Dyspn: Learning dynamic affinity for image-guided depth comple- tion,

    Y. Lin, H. Yang, T. Cheng, W. Zhou, and Z. Yin, “Dyspn: Learning dynamic affinity for image-guided depth comple- tion,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 6, pp. 4596–4609, 2023

    2023

  7. [15]

    Agg- net: Attention guided gated-convolutional network for depth image completion,

    D. Chen, T. Huang, Z. Song, S. Deng, and T. Jia, “Agg- net: Attention guided gated-convolutional network for depth image completion,” inProceedings of the ieee/cvf international conference on computer vision, 2023, pp. 8853–8862

    2023

  8. [16]

    Rgb-depth fusion gan for indoor depth comple- tion,

    H. Wang, M. Wang, Z. Che, Z. Xu, X. Qiao, M. Qi, F. Feng, and J. Tang, “Rgb-depth fusion gan for indoor depth comple- tion,” inProceedings of the ieee/cvf conference on computer vision and pattern recognition, 2022, pp. 6209–6218

    2022

  9. [17]

    3d recon- struction of transparent objects exploiting surface fluorescence caused by uv irradiation,

    R. Rantoson, C. Stolz, D. Fofi, and F. M´ eriaudeau, “3d recon- struction of transparent objects exploiting surface fluorescence caused by uv irradiation,” in2010 IEEE International Con- ference on Image Processing. IEEE, 2010, pp. 2965–2968

    2010

  10. [18]

    Full 3d reconstruction of transparent objects,

    B. Wu, Y. Zhou, Y. Qian, M. Gong, and H. Huang, “Full 3d reconstruction of transparent objects,”arXiv preprint arXiv:1805.03482, 2018

    2018 arXiv

  11. [19]

    Deep depth completion of a single rgb-d image,

    Y. Zhang and T. Funkhouser, “Deep depth completion of a single rgb-d image,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 175–185

    2018

  12. [20]

    A4t: Hierarchical affordance detection for transparent objects depth reconstruc- tion and manipulation,

    J. Jiang, G. Cao, T.-T. Do, and S. Luo, “A4t: Hierarchical affordance detection for transparent objects depth reconstruc- tion and manipulation,”IEEE Robotics and Automation Let- ters, vol. 7, no. 4, pp. 9826–9833, 2022

    2022

  13. [21]

    Clear grasp: 3d shape estimation of transparent objects for manipulation,

    S. Sajjan, M. Moore, M. Pan, G. Nagaraja, J. Lee, A. Zeng, and S. Song, “Clear grasp: 3d shape estimation of transparent objects for manipulation,” in2020 IEEE international confer- ence on robotics and automation (ICRA). IEEE, 2020, pp. 3634–3642

    2020

  14. [22]

    Depthgrasp: Depth completion of transparent objects using self-attentive adversarial network with spectral residual for grasping,

    Y. Tang, J. Chen, Z. Yang, Z. Lin, Q. Li, and W. Liu, “Depthgrasp: Depth completion of transparent objects using self-attentive adversarial network with spectral residual for grasping,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 20...

    2021

  15. [23]

    Fdct: Fast depth completion for transparent objects,

    T. Li, Z. Chen, H. Liu, and C. Wang, “Fdct: Fast depth completion for transparent objects,”IEEE Robotics and Au- tomation Letters, vol. 8, no. 9, pp. 5823–5830, 2023

    2023

  16. [24]

    Transcg: A large-scale real-world dataset for transparent object depth completion and a grasping baseline,

    H. Fang, H.-S. Fang, S. Xu, and C. Lu, “Transcg: A large-scale real-world dataset for transparent object depth completion and a grasping baseline,”IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7383–7390, 2022

    2022

  17. [25]

    Tdcnet: Transparent objects depth completion with cnn- transformer dual-branch parallel network,

    X. Fan, C. Ye, A. Deng, X. Wu, M. Pan, and H. Yang, “Tdcnet: Transparent objects depth completion with cnn- transformer dual-branch parallel network,”arXiv preprint arXiv:2412.14961, 2024

    2024 arXiv

  18. [26]

    Tcr- net: Transparent object depth completion with cascade re- finements,

    D.-H. Zhai, S. Yu, W. Wang, Y. Guan, and Y. Xia, “Tcr- net: Transparent object depth completion with cascade re- finements,”IEEE Transactions on Automation Science and Engineering, 2024

    2024

  19. [27]

    An image is worth 16x16 words: Transformers for image recognition at scale,

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

    2010 arXiv

  20. [28]

    Localvit: Analyzing locality in vision trans- formers,

    Y. Li, K. Zhang, J. Cao, R. Timofte, M. Magno, L. Benini, and L. Van Goo, “Localvit: Analyzing locality in vision trans- formers,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 9598–9605

    2023

  21. [29]

    Conformer: Local features coupling global representations for visual recognition,

    Z. Peng, W. Huang, S. Gu, L. Xie, Y. Wang, J. Jiao, and Q. Ye, “Conformer: Local features coupling global representations for visual recognition,” inProceedings of the IEEE/CVF interna- tional conference on computer vision, 2021, pp. 367–376

    2021

  22. [30]

    Transunet: Transformers make strong encoders for medical image segmentation,

    J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang, L. Lu, A. L. Yuille, and Y. Zhou, “Transunet: Transformers make strong encoders for medical image segmentation,”arXiv preprint arXiv:2102.04306, 2021

    2021 arXiv

  23. [31]

    Cfformer: Cross cnn-transformer channel attention and spatial feature fusion for improved segmentation of low quality medical images,

    J. Li, Q. Xu, X. He, Z. Liu, D. Zhang, R. Wang, R. Qu, and G. Qiu, “Cfformer: Cross cnn-transformer channel attention and spatial feature fusion for improved segmentation of low quality medical images,”arXiv preprint arXiv:2501.03629, 2025

    2025 arXiv

  24. [32]

    A multilevel multimodal fusion transformer for remote sensing semantic segmentation,

    X. Ma, X. Zhang, M.-O. Pun, and M. Liu, “A multilevel multimodal fusion transformer for remote sensing semantic segmentation,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  25. [33]

    Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation,

    Z. Xing, T. Ye, Y. Yang, G. Liu, and L. Zhu, “Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 578–588

    2024

  26. [34]

    U-mamba: Enhancing long- range dependency for biomedical image segmentation,

    J. Ma, F. Li, and B. Wang, “U-mamba: Enhancing long- range dependency for biomedical image segmentation,”arXiv preprint arXiv:2401.04722, 2024

    2024 arXiv

  27. [35]

    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,” Advances in neural information processing systems, vol. 37, pp. 103 031–103 063, 2024

    2024

  28. [36]

    Mambavision: A hy- brid mamba-transformer vision backbone,

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

    2024 arXiv

  29. [37]

    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,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 6, 2025, pp. 6443–6451

    2025

  30. [38]

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

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

    2024 arXiv

  31. [39]

    Mamba- unet: Unet-like pure visual mamba for medical image segmen- tation,

    Z. Wang, J.-Q. Zheng, Y. Zhang, G. Cui, and L. Li, “Mamba- unet: Unet-like pure visual mamba for medical image segmen- tation,”arXiv preprint arXiv:2402.05079, 2024

    2024 arXiv

  32. [40]

    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,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012– 10 022

    2021

  33. [41]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  34. [42]

    Hetero-unet: Heterogeneous transformer with mamba for medical image segmentation,

    Z. Yan, Y. Liu, X. Li, and L. Sun, “Hetero-unet: Heterogeneous transformer with mamba for medical image segmentation,” in Advancements In Medical Foundation Models: Explainability, Robustness, Security, and Beyond

  35. [43]

    Rgb-d local implicit function for depth completion of transparent objects,

    L. Zhu, A. Mousavian, Y. Xiang, H. Mazhar, J. van Eenbergen, S. Debnath, and D. Fox, “Rgb-d local implicit function for depth completion of transparent objects,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4649–4658

    2021

  36. [44]

    Seeing glass: joint point cloud and depth completion for transparent objects,

    H. Xu, Y. R. Wang, S. Eppel, A. Aspuru-Guzik, F. Shkurti, and A. Garg, “Seeing glass: joint point cloud and depth completion for transparent objects,”arXiv preprint arXiv:2110.00087, 2021

    2021 arXiv

  37. [45]

    Tode- trans: Transparent object depth estimation with transformer,

    K. Chen, S. Wang, B. Xia, D. Li, Z. Kan, and B. Li, “Tode- trans: Transparent object depth estimation with transformer,” in2023 IEEE international conference on robotics and au- tomation (ICRA). IEEE, 2023, pp. 4880–4886

    2023

  38. [46]

    Transparent depth com- pletion using segmentation features,

    B. Liu, H. Li, Z. Wang, and T. Xue, “Transparent depth com- pletion using segmentation features,”ACM Transactions on Multimedia Computing, Communications and Applications, vol. 20, no. 12, pp. 1–19, 2024

    2024

Pith tools

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