Pith. sign in

REVIEW 5 major objections 5 minor 36 references

Geo-RepNet: Geometry-Aware Representation Learning for Surgical Phase Recognition in Endoscopic Submucosal Dissection

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Geo-RepNet claims that injecting depth-derived geometric priors into a re-parameterizable convolutional backbone substantially improves surgical phase recognition in endoscopic submucosal dissection (ESD), reaching 85.02% accuracy versus…

desk verdict First depth-based surgical phase recognition paper with a coherent internal ablation, but the missing depth-source and split details keep it conditional. read the letter →

arxiv 2507.09294 v1 pith:WKAXVI2B submitted 2025-07-12 cs.CV cs.RO

Rui Tang , Haochen Yin , Guankun Wang , Long Bai , An Wang , Huxin Gao , Jiazheng Wang , Hongliang Ren This is my paper
classification cs.CVcs.RO
keywords surgicalphaserecognitionendoscopicsubmucosaldissectiondepth-guidedpriorRepVGGgeometry-awareattentionRGB-Dfusionmulti-scalevideounderstanding
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

Geo-RepNet claims that injecting depth-derived geometric priors into a re-parameterizable convolutional backbone substantially improves surgical phase recognition in endoscopic submucosal dissection (ESD). On a new nine-phase ESD dataset, the method reaches 85.02% accuracy, 81.74% F1, and 93.10% AUC, versus 77.53% accuracy for the RGB-only baseline, beating both convolutional and transformer backbones. The paper argues that depth, used as a spatial attention guide rather than a second input stream, disambiguates visually similar phases and compensates for low-texture regions. The ablation study shows every component contributes, with the full model outperforming all partial combinations.

What carries the argument

The Depth-Guided Geometric Prior Generation (DGPG) module and the Geometry-Enhanced Multi-scale Attention (GEMA) block are the central mechanisms. DGPG turns a raw depth map into two priors—a sinusoid-encoded relative positional basis and a decay mask blending coordinate- and depth-based distances—so depth acts as a spatial attention guide instead of an explicit feature stream. GEMA injects these priors into attention through geometry-aware cross self-attention and efficient multi-scale aggregation, positioned after the first stage of a RepVGG backbone. This carries the argument by showing that geometric structure, not additional depth features, drives the accuracy gain.

What would settle it

Retrain the model with the depth channel replaced by random noise or by depth maps from a different video while keeping RGB frames fixed: if accuracy stays near 85%, the geometric prior carries no information, and re-running the split with videos kept intact tests whether video-level leakage accounts for the gain.

Watch

Extended reading notes

Core claim

The central claim is that raw depth maps, converted into lightweight positional and visibility priors, give a convolutional network structural cues that appearance alone lacks for distinguishing surgical phases that look alike. Geo-RepNet implements this by inserting the Depth-Guided Geometric Prior Generation (DGPG) module after the first RepVGG stage, which computes relative depth differences and sinusoidal positional encodings, and the Geometry-Enhanced Multi-scale Attention (GEMA) block, which uses these priors to modulate cross-attention and multi-scale aggregation. The authors demonstrate the point with ablations: removing all depth and attention modules drops accuracy to 77.53%, while the full model reaches 85.02%, and each added component raises F1 and AUC. The paper further claims the design stays efficient and robust, with even the largest grouping factor outperforming all baselines.

Load-bearing premise

The depth maps fed to the model are valid, spatially aligned geometric measurements rather than artifacts or hidden correlates of the RGB frames.

Editorial extensions

If this is right

  • Depth-guided geometric priors can disambiguate surgical phases that RGB appearance alone cannot separate, a prerequisite for reliable workflow monitoring and intraoperative assistance.
  • Injecting priors as attention modulators rather than a second encoder keeps computational overhead low, making real-time deployment in resource-constrained operating rooms feasible.
  • The nine-phase ESD dataset provides a new benchmark for phase recognition in endoscopic submucosal dissection, with realistic class imbalance.
  • Each component of the method contributes independently and synergistically, so the design can be partially reused in other surgical vision tasks.

Reading between the lines

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

  • If the depth maps are produced by a monocular estimation network, part of the gain may reflect the estimator's priors about endoscopic scenes rather than true geometry; replacing them with hardware-sensed depth would separate these factors.
  • The same recipe—depth as a spatial attention prior rather than a second input stream—could transfer to other low-texture endoscopic procedures such as colonoscopy or gastroscopy.
  • The relative positional encodings and visibility masks are generic enough to apply to other fine-grained surgical action or workflow recognition tasks beyond phase classification.
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

5 major / 5 minor

Summary. Geo-RepNet is a convolutional framework for surgical phase recognition in endoscopic submucosal dissection (ESD). It takes an RGB frame and a corresponding depth map, uses a RepVGG backbone, and augments it with a Depth-Guided Geometric Prior Generation (DGPG) module that converts depth into positional encodings and visibility masks, together with a Geometry-Enhanced Multi-scale Attention (GEMA) block that injects these priors. The authors construct a nine-phase ESD video dataset with 7080 training and 792 validation frames and report 85.02% accuracy, 81.74% F1, and 93.10% AUC, outperforming ten generic classification backbones on their private split. An ablation attributes most of the gain to DGPG, the geometry-aware cross self-attention (GSA), and the multi-scale attention module.

Significance. If the reported results are reliable, this is a useful first demonstration that depth-derived geometric priors can improve phase recognition in ESD, and the re-parameterizable design is attractive for real-time deployment. The paper has clear strengths: a concrete formulation of geometry priors, a coherent ablation narrative, use of real clinical ESD video data, and an efficiency-oriented architecture. The significance is capped by three empirical gaps: the depth source is undisclosed, the train/test split is not shown to be video-disjoint, and all reported numbers are single runs on a private dataset. These gaps make the headline 7.5-point gain over the RGB-only baseline difficult to interpret as evidence for the specific claim that true geometric structure is the cause.

major comments (5)
  1. [II.A, Eq. (1), and III.A] The paper never states how the depth maps d are obtained. If d is produced by a monocular depth estimation network, then d is a deterministic function of the RGB input, and the DGPG/GEMA modules cannot introduce genuinely new scene information; the 85.02% versus 77.53% gain could instead reflect the estimator's pretraining priors or hidden appearance cues rather than true geometry. Please state the depth sensor or estimation method, the network that produced the maps, its training data, and the alignment/preprocessing pipeline. If monocular depth is used, the central claim requires a controlled comparison (e.g., the same estimator's output as an RGB-derived auxiliary branch) or a carefully worded interpretation that does not assert independent geometric signal.
  2. [III.A / Table I] The train/test split is described only by per-class sample counts; there is no statement that videos are disjoint between the two splits. Endoscopic videos exhibit strong temporal autocorrelation, so random frame-level splitting can place near-duplicate frames from the same procedure in both training and test, inflating all reported metrics. Because every ablation uses the same split, the ablation cannot rule out leakage. Please specify video-level splitting, the number of videos in each split, and report per-video or leave-one-video-out performance.
  3. [Tables II-IV / Section III.C] All metrics are from a single run with no standard deviation, confidence interval, or significance test, on a test set of only 792 frames with heavy class imbalance (e.g., 9 versus 294 samples per class in Table I). Differences such as 84.77 versus 85.02 (Table IV) and 82.40 versus 85.02 (Table III) may be within run-to-run noise. Please report multiple seeds, error bars, and statistical significance, particularly for the ablation claims that attribute specific gains to individual modules.
  4. [III.C.1 / Table II] The claim of state-of-the-art performance is not supported by the current comparison set: all baselines are generic image classification backbones, with no surgical-phase-recognition methods such as TeCNO, Trans-SVNet, or temporal convolutional/transformer models, and the evaluation dataset is private. To support the headline claim, please compare against published surgical phase recognition methods on the same data and/or evaluate on an existing public benchmark such as Cholec80.
  5. [II.C / III.C.2 / Tables III-IV] The ablation uses 'EMMA (Efficient Multi-scale Modality-aware Attention)' while the architecture section defines 'EMA (Efficient Multi-scale Attention)' and Table IV again says 'EMA.' If EMMA is meant to be EMA, the naming should be unified; if it is a different module, it is never formally defined. The current text makes it unclear which component is being ablated, which weakens the attribution of the accuracy gain to the proposed modules.
minor comments (5)
  1. [II.B, Eq. (1)] Eq. (1) promises that P(·) is defined in Section II-B, but that section never explicitly defines the outputs rsin, rcos, and mvis in terms of the inputs; it introduces PEsin, PEcos, and Mgeo separately. Please add explicit definitions or clarify the notation.
  2. [Throughout] The text alternates between 'EMMA,' 'EMA,' and 'GEMA' inconsistently; please standardize nomenclature so that the module and its sub-components are unambiguous.
  3. [III.B] Implementation details list only optimizer, epochs, batch size, and learning rate; they omit input resolution, data augmentation, class-imbalance handling, and the parameter/FLOPs counts that would substantiate the computational-efficiency claims.
  4. [Table I and Fig. 1] Several phase names are awkward or grammatically incomplete (e.g., 'Forceps and Knife move to the edge mucosa flap'); the labels should be cleaned and made consistent with the captions in Figure 1.
  5. [Eq. (3)] The decay factor in Eq. (3) is not intuitively explained; please specify its valid range and clarify whether larger decay corresponds to weaker or stronger attention, since Mgeo in Eq. (5) is a weighted combination of masks.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported gains are measured predictions, not inputs renamed as results; self-citations are contextual only.

full rationale

The central claim (Tables II and III) is an empirical comparison on a held-out test split; accuracy, F1, and AUC are measured outputs of a trained network, not quantities reconstructed from the inputs by construction. The geometric priors are computed from the input depth map via Eqs. (1)-(5) and injected through attention in Eqs. (6)-(12); no equation defines the phase label as a function of the prior in a way that makes the reported improvement tautological. The references include several self-citations (e.g., [1], [2], [12], [16]), but these are used for background, dataset context, or depth-estimation motivation, not to justify the classification result or to forbid alternative architectures. The undisclosed provenance of the depth maps and the unknown granularity of the train/test split are correctness and validity risks, not circularity: if depth were generated by a monocular network trained on the same videos, the claimed gain could be reinterpreted or confounded, but that would be a data-provenance or leakage concern rather than the prediction being identical to a fitted parameter or restating an input by definition. Thus no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The method is a learned system whose millions of weights are fitted to the training split, which is standard practice. The entries above are the hand-chosen hyperparameters and unstated premises the reported performance depends on. No entirely novel physical or scientific entities are introduced; the depth prior is an architectural construct whose independent evidential support is confined to the authors' own ablation.

free parameters (4)
  • EMA grouping factor = 4
    Selected as the best setting in Table IV; the paper reports accuracy drops as the factor grows to 32, so this choice affects the reported headline numbers.
  • DGPG decay parameters lambda_0 and gamma = not reported
    Learnable parameters in Eq. (3) controlling the head-specific exponential depth decay; initialization and ranges are not stated.
  • Depth-visibility fusion weights w1 and w2 = not reported
    Learnable weights in Eq. (5) balancing positional and depth-based decay masks.
  • Prior injection location = after Stage 1 of RepVGG
    Chosen by hand with a qualitative justification; no ablation across stages is reported.
assumptions (4)
  • domain assumption Depth maps are available for every training and test frame and are aligned with the RGB frames.
    Invoked implicitly in Eq. (1) and Section II.A; the acquisition or estimation pipeline is never described, so this premise is unverified.
  • domain assumption Train and validation frames do not share video-level provenance.
    Section III.A describes the split into 7080 and 792 frames but does not state whether the split is video-disjoint; frame correlation across splits would inflate all metrics.
  • domain assumption The nine hand-selected phase labels are clinically coherent and the dense annotations are consistent.
    Section III.A; annotation protocol and inter-annotator agreement are not reported, and some phases have very few frames (e.g., Forceps rotate with 3 test frames).
  • standard math Standard stochastic training with Adam and cosine annealing converges to a representative solution.
    Section III.B; no seeds, runs, or variance are reported, so the single-run numbers are treated as point estimates.
invented entities (1)
  • Depth-Guided Geometric Prior (relative positional bases and visibility decay mask)
    purpose: Condenses raw depth into angular positional embeddings and an attention decay mask that steers the network toward structurally informative regions.
    Its benefit is demonstrated only through the internal ablation on the private dataset; there is no external benchmark or public artifact to confirm the prior behaves as described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geo-RepNet: Geometry-Aware Representation Learning for Surgical Phase Recognition in Endoscopic Submucosal Dissection." pith.science (2026). https://pith.science/paper/WKAXVI2B

@misc{pith2026250709294,
  author       = {Pith},
  title        = {Pith review of: Geo-RepNet: Geometry-Aware Representation Learning for Surgical Phase Recognition in Endoscopic Submucosal Dissection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WKAXVI2B}},
  note         = {Machine review of arXiv:2507.09294}
}
read the original abstract

Surgical phase recognition plays a critical role in developing intelligent assistance systems for minimally invasive procedures such as Endoscopic Submucosal Dissection (ESD). However, the high visual similarity across different phases and the lack of structural cues in RGB images pose significant challenges. Depth information offers valuable geometric cues that can complement appearance features by providing insights into spatial relationships and anatomical structures. In this paper, we pioneer the use of depth information for surgical phase recognition and propose Geo-RepNet, a geometry-aware convolutional framework that integrates RGB image and depth information to enhance recognition performance in complex surgical scenes. Built upon a re-parameterizable RepVGG backbone, Geo-RepNet incorporates the Depth-Guided Geometric Prior Generation (DGPG) module that extracts geometry priors from raw depth maps, and the Geometry-Enhanced Multi-scale Attention (GEMA) to inject spatial guidance through geometry-aware cross-attention and efficient multi-scale aggregation. To evaluate the effectiveness of our approach, we construct a nine-phase ESD dataset with dense frame-level annotations from real-world ESD videos. Extensive experiments on the proposed dataset demonstrate that Geo-RepNet achieves state-of-the-art performance while maintaining robustness and high computational efficiency under complex and low-texture surgical environments.

Figures

Figures reproduced from arXiv: 2507.09294 by the authors.

Figure 1
Figure 1. Different surgical phases can exhibit high visual feature similarity, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our Geo-RepNet framework. (a) Architecture of Geo-RepNet; (b) Depth-Guided Geometric Prior Generation Module; (c) Geometry [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 24 canonical work pages

  1. [1]

    Copesd: A multi-level surgical motion dataset for training large vision-language models to co-pilot endoscopic submucosal dissection,

    G. Wang, H. Xiao, H. Gao, R. Zhang, L. Bai, X. Yang, Z. Li, H. Li, and H. Ren, “Copesd: A multi-level surgical motion dataset for training large vision-language models to co-pilot endoscopic submucosal dissection,” arXiv preprint arXiv:2410.07540 , 2024

    arXiv

  2. [2]

    Ossar: Towards open-set surgical activity recognition in robot-assisted surgery,

    L. Bai, G. Wang, J. Wang, X. Yang, H. Gao, X. Liang, A. Wang, M. Is- lam, and H. Ren, “Ossar: Towards open-set surgical activity recognition in robot-assisted surgery,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 14 622–14 629

    work page 2024

  3. [3]

    Tecno: Surgical phase recognition with multi- stage temporal convolutional networks,

    T. Czempiel, M. Paschali, M. Keicher, W. Simson, H. Feussner, S. T. Kim, and N. Navab, “Tecno: Surgical phase recognition with multi- stage temporal convolutional networks,” in Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part III 23 . Springer, 2020, pp. 343–352

    work page 2020

  4. [4]

    Multimodal graph representation learning for robust surgical workflow recognition with adversarial feature disentanglement,

    L. Bai, B. Ma, R. Wang, G. Wang, B. Cui, Z. Jiang, M. Islam, Z. Min, J. Lai, N. Navab et al. , “Multimodal graph representation learning for robust surgical workflow recognition with adversarial feature disentanglement,” Information Fusion, p. 103290, 2025

    work page 2025

  5. [5]

    Trans-svnet: Accurate phase recognition from surgical videos via hybrid embedding aggregation transformer,

    X. Gao, Y . Jin, Y . Long, Q. Dou, and P.-A. Heng, “Trans-svnet: Accurate phase recognition from surgical videos via hybrid embedding aggregation transformer,” in Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part IV 24. Springer, 2021, ...

    work page 2021

  6. [6]

    Multi-task temporal convo- lutional networks for joint recognition of surgical phases and steps in gastric bypass procedures,

    S. Ramesh, D. Dall’Alba, C. Gonzalez, T. Yu, P. Mascagni, D. Mutter, J. Marescaux, P. Fiorini, and N. Padoy, “Multi-task temporal convo- lutional networks for joint recognition of surgical phases and steps in gastric bypass procedures,” International journal of computer assisted radiology and surgery , vol. 16, pp. 1111–1119, 2021

    work page 2021

  7. [7]

    Asformer: Transformer for action segmen- tation,

    F. Yi, H. Wen, and T. Jiang, “Asformer: Transformer for action segmen- tation,” arXiv preprint arXiv:2110.08568 , 2021

    arXiv 2021

  8. [8]

    Training data-efficient image transformers & distillation through attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. J ´egou, “Training data-efficient image transformers & distillation through attention,” in International conference on machine learning . PMLR, 2021, pp. 10 347–10 357

    2021

Show all 36 references
  1. [9]

    Recognizing surgical phases anywhere: Few-shot test-time adaptation and task-graph guided refinement,

    K. Yuan, T. Chen, S. Li, J. L. Lavanchy, C. Heiliger, E. ¨Ozsoy, Y . Huang, L. Bai, N. Navab, V . Srivastav et al., “Recognizing surgical phases anywhere: Few-shot test-time adaptation and task-graph guided refinement,” arXiv preprint arXiv:2506.20254 , 2025

    2025 arXiv

  2. [10]

    Repvgg: Mak- ing vgg-style convnets great again,

    X. Ding, X. Zhang, N. Ma, J. Han, G. Ding, and J. Sun, “Repvgg: Mak- ing vgg-style convnets great again,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 13 733–13 742

    2021

  3. [11]

    A survey on deep learning in medical image analysis,

    G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A. Van Der Laak, B. Van Ginneken, and C. I. S´anchez, “A survey on deep learning in medical image analysis,” Medical image analysis, vol. 42, pp. 60–88, 2017

    2017

  4. [12]

    Endodac: Efficient adapting foundation model for self-supervised depth estimation from any endoscopic camera,

    B. Cui, M. Islam, L. Bai, A. Wang, and H. Ren, “Endodac: Efficient adapting foundation model for self-supervised depth estimation from any endoscopic camera,” in International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 2024, pp. 208–218

    2024

  5. [13]

    Bi-directional cross-modality feature propagation with separation-and- aggregation gate for rgb-d semantic segmentation,

    X. Chen, K.-Y . Lin, J. Wang, W. Wu, C. Qian, H. Li, and G. Zeng, “Bi-directional cross-modality feature propagation with separation-and- aggregation gate for rgb-d semantic segmentation,” in European confer- ence on computer vision . Springer, 2020, pp. 561–577

    2020

  6. [14]

    Prediction of laparoscopic procedure duration using unlabeled, multimodal sen- sor data,

    S. Bodenstedt, M. Wagner, L. M ¨undermann, H. Kenngott, B. M ¨uller- Stich, M. Breucha, S. T. Mees, J. Weitz, and S. Speidel, “Prediction of laparoscopic procedure duration using unlabeled, multimodal sen- sor data,” International Journal of Computer Assisted Radiology and Sur...

    2019

  7. [15]

    V 2-sfmlearner: Learning monocular depth and ego-motion for multimodal wireless capsule endoscopy,

    L. Bai, B. Cui, L. Wang, Y . Li, S. Yao, S. Yuan, Y . Wu, Y . Zhang, M. Q.-H. Meng, Z. Li et al. , “V 2-sfmlearner: Learning monocular depth and ego-motion for multimodal wireless capsule endoscopy,”IEEE Transactions on Automation Science and Engineering , 2025

    2025

  8. [16]

    Surgical-dino: adapter learning of foundation models for depth estimation in endoscopic surgery,

    B. Cui, M. Islam, L. Bai, and H. Ren, “Surgical-dino: adapter learning of foundation models for depth estimation in endoscopic surgery,” International Journal of Computer Assisted Radiology and Surgery , vol. 19, no. 6, pp. 1013–1020, 2024

    2024

  9. [17]

    Dense depth estimation in monocular endoscopy with self-supervised learning methods,

    X. Liu, A. Sinha, M. Ishii, G. D. Hager, A. Reiter, R. H. Taylor, and M. Unberath, “Dense depth estimation in monocular endoscopy with self-supervised learning methods,” IEEE transactions on medical imaging, vol. 39, no. 5, pp. 1438–1447, 2019

    2019

  10. [18]

    A survey on indoor rgb-d semantic segmentation: from hand-crafted features to deep convolutional neural networks,

    F. Fooladgar and S. Kasaei, “A survey on indoor rgb-d semantic segmentation: from hand-crafted features to deep convolutional neural networks,” Multimedia Tools and Applications, vol. 79, no. 7, pp. 4499– 4524, 2020

    2020

  11. [19]

    Token pooling in vision transformers,

    D. Marin, J.-H. R. Chang, A. Ranjan, A. Prabhu, M. Rastegari, and O. Tuzel, “Token pooling in vision transformers,” arXiv preprint arXiv:2110.03860, 2021

    2021 arXiv

  12. [20]

    Deep learning in medical image analysis,

    D. Shen, G. Wu, and H.-I. Suk, “Deep learning in medical image analysis,” Annual review of biomedical engineering , vol. 19, no. 1, pp. 221–248, 2017

    2017

  13. [21]

    Stand-alone self-attention in vision models,

    P. Ramachandran, N. Parmar, A. Vaswani, I. Bello, A. Levskaya, and J. Shlens, “Stand-alone self-attention in vision models,” Advances in neural information processing systems , vol. 32, 2019

    2019

  14. [22]

    Non-local neural net- works,

    X. Wang, R. Girshick, A. Gupta, and K. He, “Non-local neural net- works,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7794–7803

    2018

  15. [23]

    Pvt v2: Improved baselines with pyramid vision transformer,

    W. Wang, E. Xie, X. Li, D.-P. Fan, K. Song, D. Liang, T. Lu, P. Luo, and L. Shao, “Pvt v2: Improved baselines with pyramid vision transformer,” Computational visual media , vol. 8, no. 3, pp. 415–424, 2022

    2022

  16. [24]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022

    2021

  17. [25]

    Dformerv2: Geometry self-attention for rgbd semantic segmentation,

    B.-W. Yin, J.-L. Cao, M.-M. Cheng, and Q. Hou, “Dformerv2: Geometry self-attention for rgbd semantic segmentation,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 19 345– 19 355

    2025

  18. [26]

    Efficient multi-scale attention module with cross-spatial learning,

    D. Ouyang, S. He, G. Zhang, M. Luo, H. Guo, J. Zhan, and Z. Huang, “Efficient multi-scale attention module with cross-spatial learning,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

    2023

  19. [27]

    Mobilevitv3: Mobile-friendly vision transformer with simple and effective fusion of local, global and input features,

    S. N. Wadekar and A. Chaurasia, “Mobilevitv3: Mobile-friendly vision transformer with simple and effective fusion of local, global and input features,” arXiv preprint arXiv:2209.15159 , 2022

    2022 arXiv

  20. [28]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 11 976–11 986

    2022

  21. [29]

    Run, don’t walk: chasing higher flops for faster neural networks,

    J. Chen, S.-h. Kao, H. He, W. Zhuo, S. Wen, C.-H. Lee, and S.-H. G. Chan, “Run, don’t walk: chasing higher flops for faster neural networks,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 12 021–12 031

    2023

  22. [30]

    Res2net: A new multi-scale backbone architecture,

    S.-H. Gao, M.-M. Cheng, K. Zhao, X.-Y . Zhang, M.-H. Yang, and P. Torr, “Res2net: A new multi-scale backbone architecture,” IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 2, pp. 652–662, 2019

    2019

  23. [31]

    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

  24. [32]

    Swin transformer v2: Scaling up capacity and resolution,

    Z. Liu, H. Hu, Y . Lin, Z. Yao, Z. Xie, Y . Wei, J. Ning, Y . Cao, Z. Zhang, L. Dong et al., “Swin transformer v2: Scaling up capacity and resolution,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 12 009–12 019

    2022

  25. [33]

    Twins: Revisiting the design of spatial attention in vision trans- formers,

    X. Chu, Z. Tian, Y . Wang, B. Zhang, H. Ren, X. Wei, H. Xia, and C. Shen, “Twins: Revisiting the design of spatial attention in vision trans- formers,” Advances in neural information processing systems , vol. 34, pp. 9355–9366, 2021

    2021

  26. [34]

    Uniformer: Unifying convolution and self-attention for visual recogni- tion,

    K. Li, Y . Wang, J. Zhang, P. Gao, G. Song, Y . Liu, H. Li, and Y . Qiao, “Uniformer: Unifying convolution and self-attention for visual recogni- tion,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 10, pp. 12 581–12 600, 2023

    2023

  27. [35]

    Cswin transformer: A general vision transformer backbone with cross-shaped windows,

    X. Dong, J. Bao, D. Chen, W. Zhang, N. Yu, L. Yuan, D. Chen, and B. Guo, “Cswin transformer: A general vision transformer backbone with cross-shaped windows,” in Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition , 2022, pp. 12 124– 12 134

    2022

  28. [36]

    Mambaout: Do we really need mamba for vision?

    W. Yu and X. Wang, “Mambaout: Do we really need mamba for vision?” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 4484–4496

    2025

Pith tools

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