Pith. sign in

REVIEW 2 major objections 5 minor 55 references

EMRA-proxy: Enhancing Multi-Class Region Semantic Segmentation in Remote Sensing Images with Attention Proxy

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A region-level attention proxy outperforms pixel-grid segmentation on remote sensing images.

desk verdict Useful incremental architecture, but the LoveDA SOTA comparison is contaminated by a self-ablation row passed off as Segmenter; fix before trusting the numbers. read the letter →

arxiv 2505.17665 v1 pith:R3TYENNO submitted 2025-05-23 cs.CV cs.AI

classification cs.CVcs.AI
keywords semanticsegmentationremotesensingregionproxymulti-classattentionVisionTransformerlandcoverHRA-proxyMCA-proxy
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 claims that semantic segmentation of high-resolution remote sensing images is better done at the region level than at the pixel level. It introduces EMRA-proxy, which uses a Vision Transformer to encode learnable region proxies, a Homogeneous Regions Attention Proxy (HRA-proxy) to capture region geometry through local pixel-to-region associations, and a Multi-class Attention Proxy (MCA-proxy) to extract category information from extra class tokens. The two output maps are fused to produce per-pixel predictions. On LoveDA, Potsdam, and Vaihingen, the method reports mIoU of 55.22%, 74.39%, and 70.03%, which the authors state are above previous state-of-the-art results.

What carries the argument

The key mechanism is the region proxy: every token in the ViT encoder represents a learnable region, and each pixel is associated with tokens in a fixed 3x3 neighborhood through a local affinity graph that forms a tessellation of the image plane. HRA-proxy uses the first M transformer layers and a depthwise convolution to produce a Homogeneous Semantic Mask Feature Map (HSMF-map). MCA-proxy prepends C class tokens, aggregates attention from the last P layers and multiple heads, and produces a Global Class Attention Map (GCA-map). The final per-pixel logits are computed as a sum over neighboring regions of the class logits weighted by the pixel-region association probabilities.

What would settle it

An independent implementation with the same training schedule should reproduce the reported mIoU values (55.22% on LoveDA, 74.39% on Potsdam, 70.03% on Vaihingen); if the model with the fixed 3x3 association does not match those numbers, or if changing the neighborhood to 5x5 changes results materially, the central claim of the paper is unsupported.

Watch

Extended reading notes

Core claim

The central discovery is that a model operating purely on region embeddings, with no pixel-grid decoder, can outperform both CNN- and Transformer-based segmenters on multi-class remote sensing images. The method's accuracy comes from combining two complementary proxies: HRA-proxy, which learns a homogeneous semantic mask feature map from early encoder layers, and MCA-proxy, which builds a global class attention map from the last encoder layers. Their fusion yields sharper boundaries and better handling of categories with high intra-class variation. The paper reports that the full model improves LoveDA mIoU from 38.15% for a bare ViT-B/16 baseline to 55.22%.

Load-bearing premise

The method assumes that associating every pixel with a fixed 3x3 set of neighboring regions is enough to capture object boundaries; the paper does not vary this neighborhood size in its experiments, so the choice is untested on the target datasets.

Editorial extensions

If this is right

  • Semantic segmentation can be performed without a dense pixel-grid decoder; a linear classifier on region embeddings plus a lightweight association map is enough.
  • Adding extra class tokens and aggregating multi-layer attention improves multi-class localization for remote sensing categories with large intra-class variation.
  • The method reports consistent gains over CNN and Transformer baselines on three datasets with different object scales and urban/rural content.
  • The combination of HRA-proxy and MCA-proxy is complementary: HRA-proxy handles complex geometric shapes, while MCA-proxy improves category-level localization.
  • Larger ViT backbones improve accuracy, with ViT-B offering the best trade-off between mIoU and computational cost.

Reading between the lines

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

  • If the reported gains replicate, region-proxy decoding could be applied to other dense prediction tasks such as change detection or panoptic segmentation of remote sensing imagery, where boundary coherence matters.
  • The fixed 3x3 pixel-to-region neighborhood is a natural hyperparameter to adapt to image resolution; varying it on the same datasets is a direct testable extension the paper leaves open.
  • The class-token attention mechanism could be reused for weakly supervised segmentation, where only image-level labels are available, an application the paper does not discuss.
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

2 major / 5 minor

Summary. The paper proposes EMRA-proxy, a remote sensing semantic segmentation model that operates at the region level rather than the pixel-grid level. It combines a Homogeneous Regions Attention Proxy (HRA-proxy), which learns region embeddings and produces a homogeneous semantic mask feature map via local pixel-region associations, with a Multi-class Attention Proxy (MCA-proxy), which extracts global class attention maps from additional class tokens. The two maps are fused to produce the final segmentation. The method is evaluated on LoveDA, Potsdam, and Vaihingen, reporting state-of-the-art mIoU values of 55.22%, 74.39%, and 70.03%, respectively. The paper includes ablations on backbone size, depths M and P, and shows qualitative improvements.

Significance. If the reported results are valid, the region-proxy design is a plausible and useful direction for high-resolution remote sensing segmentation, particularly for irregular object boundaries and multi-class scenes. The approach is not circular: the model is trained end-to-end on public benchmarks, and hyperparameters M, P, and neighborhood size are chosen via validation. However, the LoveDA comparison table contains a baseline row that is numerically identical to the authors' own MCA-only ablation, which directly undermines the headline claim of a 2.08% improvement over the state of the art on that dataset. The fixed 3x3 pixel-region neighborhood is also asserted without supporting ablation. The paper provides no code or error bars, which limits reproducibility. The core idea is interesting and the experimental protocol is otherwise standard, but the LoveDA baseline issue must be resolved before the central claims can be accepted.

major comments (2)
  1. [Section 4.3, Table 6] The row labeled 'Segmenter [37] ICCV21' in Table 6 reports per-class IoUs of (51.43, 60.4, 55.81, 69.44, 30.86, 40.66, 59.58) and an mIoU of 52.60. These numbers are identical, to the decimal, to the 'EMRA-proxy-B/16' row with only MCA-proxy enabled in Table 4 of this paper. Since MCA-proxy is presented as a novel module added to a ViT decoder, a genuine Segmenter baseline cannot coincide with this ablation row unless the module has zero effect, which contradicts the other rows in Table 4. This suggests a copy-paste error in the baseline. The claimed 2.08% mIoU improvement over the previous best method on LoveDA is therefore measured against the authors' own model variant, not against the external Segmenter baseline. Please rerun or correctly report the Segmenter baseline and update the headline comparison, the abstract, and the conclusion accordingly.
  2. [Section 3.2, Eq. (3)] The pixel-region association fixes a 3x3 neighborhood with |N_p|=9 and states, without evidence, that this is 'effective for all model sizes.' No ablation varying N_p is provided on any of the three target datasets. Because the HSMF-map and the fusion in Eq. (5) depend directly on this association, the universal-effectiveness claim is unsupported. Please provide an ablation over neighborhood sizes (e.g., 1x1, 3x3, 5x5) or temper the claim to the tested setting.
minor comments (5)
  1. [Abstract / Title] The abstract provided in the submission header describes a 'Region-Aware Proxy Network (RAPNet)' with modules 'Contextual Region Attention (CRA)' and 'Global Class Refinement (GCR)', which do not match the EMRA-proxy, HRA-proxy, and MCA-proxy presented in the body of the paper. The abstract must be aligned with the actual content of the manuscript.
  2. [Section 4.3, Table 6] The 'Ours' row in Table 6 reports an mIoU of 55.22%, which is the multi-scale value from Table 3, while the comparison baselines do not state whether they use single-scale or multi-scale inference. Please clarify whether all methods are evaluated under the same inference protocol, or report both scales explicitly.
  3. [Table 5] There is a typo in the caption: 'EMCR-proxy' should be 'EMRA-proxy'. Additionally, the table reports single-scale mIoU but does not state the backbone used; please specify that these results are for ViT-B/16.
  4. [Section 3.3] The sentence 'We conducted a series of comparative experiments to determine how to choose K, as illustrated in .TheImpactofParameterPontheAccuracyofMCA-proxy..' contains a broken reference and an incomplete sentence. The intended cross-reference to Fig. 12 should be fixed.
  5. [Reproducibility] The paper states that the method is implemented in mmsegmentation but provides no code or model weights. Given that the reported gains are modest and the LoveDA baseline is currently in question, releasing code would be important for verifying the experimental claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EMRA-proxy is trained end-to-end on external benchmarks, and no fitted parameter is presented as a prediction. A duplicated Segmenter baseline is flagged as a data-integrity issue, not a circular derivation.

full rationale

The derivation chain is self-contained with respect to circularity. HRA-proxy and MCA-proxy are defined architecturally (Eqs. 3-5) and trained end-to-end with cross-entropy loss on three public datasets; the reported test metrics are not obtained by fitting a parameter to the test set. Hyperparameters M, P, and the 3x3 neighborhood are selected on validation splits (Table 5, Fig. 12, Section 3.2) and are standard model selection rather than inputs that force the reported mIoU values. The citation to RegProxy [47] is external prior work, not a self-citation, and the combined model is evaluated against RegProxy rather than derived from it. One in-scope concern, flagged for correctness but not for circularity: Table 6 labels a row as 'Segmenter [37] ICCV21' with per-class IoUs (51.43, 60.4, 55.81, 69.44, 30.86, 40.66, 59.58) and mIoU 52.60, which exactly reproduce Table 4's 'EMRA-proxy-B/16' with MCA-proxy only. This appears to be a copy/paste duplication of the authors' own ablation and would invalidate the Segmenter comparison; however, the paper's stated 2.08% SOTA improvement is computed against I2HN (53.14), not Segmenter, so the headline claim does not reduce to the authors' own ablation by construction. The inserted RAPNet abstract at the top is inconsistent with the full-text EMRA-proxy abstract; this is an editorial inconsistency, not a circular step. Therefore no circularity is found.

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

The model introduces no new physical entities. Region proxies and class tokens are internal latent representations, not independently testable objects. The free parameters are architectural hyperparameters tuned on a validation split; the central claim rests on the domain assumption that grid-based 3x3 pixel-region associations and class-token attention are adequate for high-resolution remote sensing geometry.

free parameters (4)
  • Pixel-region neighborhood size |N_p| = 3x3, |N_p|=9
    Fixed neighborhood size for pixel-region associations, asserted from empirical observations in RegProxy and not varied for remote sensing datasets.
  • M (HRA-proxy token head depth) = 3
    Selected by ablation on LoveDA validation (Table 5); controls which early transformer layers provide region features.
  • P (MCA-proxy layer aggregation count) = 4
    Selected by experiment on LoveDA validation (Fig. 12); controls how many final transformer layers are aggregated for the class attention map.
  • Output stride (w, h) = 4, 4
    Hand-set in Section 3.3 to generate a 4x stride logits map for Ours-x/16 models.
assumptions (4)
  • domain assumption Pixels with similar low-level attributes within close proximity should have homogeneous semantic meaning.
    Lifted from superpixel literature and RegProxy in Section 3.2; it is the foundation of the HRA-proxy region representation.
  • ad hoc to paper A 3x3 neighborhood with |N_p|=9 is effective for pixel-region association for all model sizes and remote sensing inputs.
    Stated in Section 3.2 after Eq. 3 based on empirical observations in RegProxy; no ablation is provided for the target datasets.
  • domain assumption Aggregating self-attention maps from the last P layers and all heads yields class-specific region logits.
    Implicit in Section 3.3; the mapping from attention to class logits is trained end to end but is not derived or independently validated.
  • domain assumption ImageNet21k pre-trained ViT weights transfer to high-resolution remote sensing images.
    Section 4.2 uses ImageNet21k pre-training following prior work; no analysis of domain shift is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EMRA-proxy: Enhancing Multi-Class Region Semantic Segmentation in Remote Sensing Images with Attention Proxy." pith.science (2026). https://pith.science/paper/R3TYENNO

@misc{pith2026250517665,
  author       = {Pith},
  title        = {Pith review of: EMRA-proxy: Enhancing Multi-Class Region Semantic Segmentation in Remote Sensing Images with Attention Proxy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R3TYENNO}},
  note         = {Machine review of arXiv:2505.17665}
}
read the original abstract

High-resolution remote sensing (HRRS) image segmentation is challenging due to complex spatial layouts and diverse object appearances. While CNNs excel at capturing local features, they struggle with long-range dependencies, whereas Transformers can model global context but often neglect local details and are computationally expensive.We propose a novel approach, Region-Aware Proxy Network (RAPNet), which consists of two components: Contextual Region Attention (CRA) and Global Class Refinement (GCR). Unlike traditional methods that rely on grid-based layouts, RAPNet operates at the region level for more flexible segmentation. The CRA module uses a Transformer to capture region-level contextual dependencies, generating a Semantic Region Mask (SRM). The GCR module learns a global class attention map to refine multi-class information, combining the SRM and attention map for accurate segmentation.Experiments on three public datasets show that RAPNet outperforms state-of-the-art methods, achieving superior multi-class segmentation accuracy.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 53 canonical work pages

  1. [37]

    G., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation

    Strudel, R., Pinel, R. G., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021,2021,pp.7242–7252

  2. [1]

    River segmentation for flood monitoring,

    L. Lopez-Fuentes, C. Rossi, and H. Skinnemoen, “River segmentation for flood monitoring,” in 2017 IEEE international conference on big data (Big Data). IEEE, 2017, pp.3746–3749

  3. [2]

    Data augmentation for deep learning based semantic segmentation and crop-weed classification in agricultural robotics,

    D. Su, H. Kong, Y. Qiao, and S. Sukkarieh, “Data augmentation for deep learning based semantic segmentation and crop-weed classification in agricultural robotics,” Computers andElectronicsinAgriculture,vol.190,p.106418,2021

  4. [3]

    Multispectral semantic land cover segmentation from aerial imagery with deep encoder–decoder network,

    C. Liu, S. Du, H. Lu, D. Li, and Z. Cao, “Multispectral semantic land cover segmentation from aerial imagery with deep encoder–decoder network,” IEEE Geoscience and Remote SensingLetters,vol.19,pp.1–5,2020

  5. [4]

    Deep learning-based remote and social sensing data fusion for urban region function recognition,

    R. Cao, W. Tu, C. Yang, Q. Li, J. Liu, J. Zhu, Q. Zhang, Q. Li, and G. Qiu, “Deep learning-based remote and social sensing data fusion for urban region function recognition,”ISPRSJournalofPhotogrammetry andRemoteSensing,vol.163,pp.82–97, 2020

  6. [5]

    Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation,

    J. Wang, Z. Zheng, A. Ma, X. Lu, and Y. Zhong, “Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation,” in Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets andBenchmarks2021,December2021,virtual,2021

  7. [6]

    Regularized least squares twin svm for multiclass classification,

    J. Ali, M. Aldhaifallah, K. S. Nisar, A. A. Aljabr, and M. Tanveer, “Regularized least squares twin svm for multiclass classification,” Big Data Research, vol. 27, p. 100295, 2022

  8. [7]

    Automatedsegmentationofirisimagesacquired in an unconstrained environment using hog-svm and growcut,

    A.Radman,N.Zainal,andS.A.Suandi,“Automatedsegmentationofirisimagesacquired in an unconstrained environment using hog-svm and growcut,” Digital Signal Processing, vol.64,pp.60–70,2017

Show all 55 references
  1. [8]

    Unsupervisedshipdetectionbasedonsaliencyand s-hog descriptor from optical satellite images,

    S.Qi,J.Ma,J.Lin,Y.Li,andJ.Tian,“Unsupervisedshipdetectionbasedonsaliencyand s-hog descriptor from optical satellite images,” IEEE geoscience and remote sensing letters,vol.12,no.7,pp.1451–1455,2015

  2. [9]

    Unsupervised domain adaptation for semantic segmentation of high-resolution remote sensing imagery,

    J. Chen, J. Zhu, Y. Guo, G. Sun, Y. Zhang, and M. Deng, “Unsupervised domain adaptation for semantic segmentation of high-resolution remote sensing imagery,” IEEE TransactionsonGeoscienceandRemoteSensing,vol.60,pp.1–15,2022. 24 YichunYuandYuqingLan

  3. [10]

    Enhanced contextual representation with deep neural networks for land cover classification based on remote sensing images,

    X. Cheng, X. He, M. Qiao, P. Li, S. Hu, P. Chang, and Z. Tian, “Enhanced contextual representation with deep neural networks for land cover classification based on remote sensing images,” International Journal of Applied Earth Observation and Geoinformation, vol.107,p.102706,2022

  4. [11]

    Object detection in high-resolution remote sensing images based on a hard-example-mining network,

    L. Zhang, Y. Wang, and Y. Huo, “Object detection in high-resolution remote sensing images based on a hard-example-mining network,” IEEE Transactions on Geoscience and RemoteSensing,vol.59,no.10,pp.8768–8780,2020

  5. [12]

    A domain adaptation neural network for change detection with heterogeneous optical and SAR remote sensing images,

    C. Zhang, Y. Feng, L. Hu, D. Tapete, L. Pan, Z. Liang, F. Cigna, and P. Yue, “A domain adaptation neural network for change detection with heterogeneous optical and SAR remote sensing images,” International Journal of Applied Earth Observation and Geoinformation,vol.109,p.102769,2022

  6. [13]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,”pp.3431–3440,2015

  7. [14]

    In: Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015 - 18th International Conference Munich, Germany, October 5 - 9, 2015, Proceedings, Part III, ser

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015 - 18th International Conference Munich, Germany, October 5 - 9, 2015, Proceedings, Part III, se...

  8. [15]

    Multistage attention ResU-Net for semantic segmentation of fine-resolution remote sensing images,

    R. Li, S. Zheng, C. Duan, J. Su, and C. Zhang, “Multistage attention ResU-Net for semantic segmentation of fine-resolution remote sensing images,” IEEE Geoscience and RemoteSensingLetters,vol.19,pp.1–5,2021

  9. [16]

    Large kernel matters–improve semantic segmentationbyglobalconvolutionalnetwork,

    C. Peng, X. Zhang, G. Yu, G. Luo, and J. Sun, “Large kernel matters–improve semantic segmentationbyglobalconvolutionalnetwork,”inProceedingsoftheIEEEconferenceon computervisionandpatternrecognition,2017,pp.4353–4361

  10. [17]

    Encoder-decoder with atrous separable convolution for semantic image segmentation,

    L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder-decoder with atrous separable convolution for semantic image segmentation,” in Proceedings of the Europeanconferenceoncomputervision(ECCV),2018,pp.801–818

  11. [18]

    Pyramid scene parsing network,

    H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing network,” in ProceedingsoftheIEEEconference oncomputer visionandpattern recognition,2017, pp. 2881–2890

  12. [19]

    Diverse region-based CNN for hyperspectral image classification,

    M. Zhang, W. Li, and Q. Du, “Diverse region-based CNN for hyperspectral image classification,” IEEE Transactions on Image Processing, vol. 27, no. 6, pp. 2623–2634, 2018

  13. [20]

    Attentionisallyouneed,

    A. Vaswani,N. Shazeer, N. Parmar,J. Uszkoreit, L. Jones,A. N.Gomez, L. Kaiser, andI. Polosukhin,“Attentionisallyouneed,”AdvancesinNeurIPS,vol.30,2017

  14. [22]

    Encoding contextual information by interlacing transformer and convolution for remote sensing imagery semanticsegmentation,

    X. Li, F. Xu, R. Xia, T. Li, Z. Chen, X. Wang, Z. Xu, and X. Lyu, “Encoding contextual information by interlacing transformer and convolution for remote sensing imagery semanticsegmentation,”RemoteSensing,vol.14,no.16,p.4065,2022

  15. [23]

    DSViT:Dynamicallyscalablevisiontransformerforremote sensingimagesegmentationandclassification,

    F.Wang,J.Ji,andY.Wang,“DSViT:Dynamicallyscalablevisiontransformerforremote sensingimagesegmentationandclassification,”IEEEJ-STARS,2023

  16. [24]

    Semantic segmentation with attention mechanism for remote sensing images,

    Q. Zhao, J. Liu, Y. Li, and H. Zhang, “Semantic segmentation with attention mechanism for remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60,pp.1–13,2021

  17. [25]

    ISPRS Journal of PhotogrammetryandRemoteSensing,vol.145,pp.78–95,2018

    Liu, Y., Fan, B., Wang, L., Bai, J., Xiang, S., Pan, C.: Semantic labeling in very high resolution images via a self-cascaded convolutional neural network. ISPRS Journal of PhotogrammetryandRemoteSensing,vol.145,pp.78–95,2018. EMRA-proxy:EnhancingMulti-ClassRegionSemanticSegme...

  18. [26]

    ISPRSJournal ofPhotogrammetryandRemoteSensing,vol.156,pp.1–13,2019

    Yue, K., Yang, L., Li, R., Hu, W., Zhang, F., Li, W.: TreeUNet: Adaptive tree convolutional neural networks forsubdecimeter aerial image segmentation. ISPRSJournal ofPhotogrammetryandRemoteSensing,vol.156,pp.1–13,2019

  19. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp

    Zheng,Z.,Zhong,Y.,Wang, J.,Ma,A.: Foreground-awarerelationnetwork forgeospatial object segmentation in high spatial resolution remote sensing imagery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 4096– 4105

  20. [28]

    Ma, A., Wang, J., Zhong, Y., Zheng, Z.: FactSeg: Foreground activation-driven small objectsemanticsegmentationinlarge-scaleremotesensingimagery.IEEETransactionson GeoscienceandRemoteSensing,vol.60,pp.1–16,2021

  21. [29]

    IEEE Transactions on Geoscience and RemoteSensing,vol.59,no.1,pp.426–435,2020

    Ding, L., Tang, H., Bruzzone, L.: Lanet: Local attention embedding to improve the semantic segmentation of remote sensing images. IEEE Transactions on Geoscience and RemoteSensing,vol.59,no.1,pp.426–435,2020

  22. [30]

    Contributors, P.: PaddleSeg, end-to-end image segmentation kit based on PaddlePaddle, 2019

  23. [31]

    In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019,2019,pp.3146–3154

    Fu,J.,Liu, J.,Tian, H.,Li, Y.,Bao,Y., Fang,Z.,Lu, H.: Dualattention network forscene segmentation. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019,2019,pp.3146–3154

  24. [32]

    Zhu, Z., Xu, M., Bai, S., Huang, T., Bai, X.: Asymmetric non-local neural networks for semanticsegmentation.In:2019IEEE/CVFInternationalConferenceonComputerVision, ICCV2019,2019,pp.593–602

  25. [33]

    CoRR, vol

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, vol. abs/2010.11929,2020

  26. [34]

    In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV2021,2021,pp.32–42

    Touvron, H., Cord, M., Sablayrolles, A., Synnaeve, G., Jégou, H.: Going deeper with image transformers. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV2021,2021,pp.32–42

  27. [35]

    In: 2021 IEEE/CVF InternationalConferenceonComputerVision,ICCV2021,2021,pp.9992–10002

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S.,Guo, B.: Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In: 2021 IEEE/CVF InternationalConferenceonComputerVision,ICCV2021,2021,pp.9992–10002

  28. [36]

    In: Computer Vision - ECCV 2020 - 16th European Conference, 2020, pp

    Yuan, Y., Chen, X., Wang, J.: Object-contextual representations for semantic segmentation. In: Computer Vision - ECCV 2020 - 16th European Conference, 2020, pp. 173–190

  29. [38]

    M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers

    Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J. M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Advances in Neural Information ProcessingSystems34:Annual ConferenceonNeuralInformationProcessing Systems2021,NeurIPS2021...

  30. [39]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V. Koltun, “Vision transformers for dense prediction,” in 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC,Canada,October10-17,2021.IEEE,2021,pp.12159–12168

  31. [40]

    Multi-class token transformer for weakly supervised semantic segmentation,

    L. Xu, W. Ouyang, M. Bennamoun, F. Boussa¨ıd, and D. Xu, “Multi-class token transformer for weakly supervised semantic segmentation,” in IEEE/CVF Conference on Computer VisionandPatternRecognition, CVPR2022, NewOrleans, LA, USA, June18- 24,2022.IEEE,2022,pp.4300–4309. 26 Yichu...

  32. [41]

    Remote sensing image change detection with transformers,

    H. Chen, Z. Qi, and Z. Shi, “Remote sensing image change detection with transformers,” IEEETransactionsonGeoscienceandRemoteSensing,vol.60,pp.1–14,2021

  33. [42]

    Pan-sharpening with customized transformer and invertible neural network,

    M. Zhou, J. Huang, Y. Fang, X. Fu, and A. Liu, “Pan-sharpening with customized transformer and invertible neural network,” in Proceedings of the AAAI conference on artificialintelligence,vol.36,no.3,2022,pp.3553–3561

  34. [43]

    Enhancing multiscale representations with transformer for remote sensing image semanticsegmentation,

    T. Xiao, Y. Liu, Y. Huang, M. Li, and G. Yang, “Enhancing multiscale representations with transformer for remote sensing image semanticsegmentation,” IEEE Transactionson GeoscienceandRemoteSensing,vol.61,pp.1–16,2023

  35. [44]

    Attention is all you need,

    A. Vaswani,N. Shazeer, N. Parmar,J. Uszkoreit, L. Jones,A. N.Gomez, L. Kaiser, andI. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December4-9,2017,LongBeach,CA...

  36. [45]

    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 Proceedings of the 38thInternationalConferenceonMachineLearning,ICML2021,18-24July2021,Virtual Event, ser. Proceedings of ...

  37. [46]

    Learning a classification model for segmentation,

    Ren and Malik, “Learning a classification model for segmentation,” in Proceedings Ninth IEEEInternationalConferenceonComputerVision,2003,pp.10–17vol.1

  38. [47]

    Semantic segmentation by early region proxy,

    Y. Zhang, B. Pang, and C. Lu, “Semantic segmentation by early region proxy,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans,LA,USA,June18-24,2022.IEEE,2022,pp.1248–1258

  39. [48]

    How to train your VIT? Data, augmentation, and regularization in vision transformers,

    A. Steiner, A. Kolesnikov, X. Zhai, R. Wightman, J. Uszkoreit, and L. Beyer, “How to train your VIT? Data, augmentation, and regularization in vision transformers,” CoRR, vol.abs/2106.10270,2021

  40. [49]

    MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark,

    M. Contributors, “MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark,”https://github.com/open-mmlab/mmsegmentation, 2020

  41. [50]

    A stochastic approximation method,

    N. K. Sinha and M. P. Griscik, “A stochastic approximation method,” IEEE Trans. Syst. ManCybern.,vol.1,no.4,pp.338–344,1971

  42. [51]

    Encoder-decoder with atrous separable convolution for semantic image segmentation,

    L. Chen et al., “Encoder-decoder with atrous separable convolution for semantic image segmentation,” in Proceedings of ECCV 2018, Munich, Germany, September 8-14, 2018, LectureNotesinComputerScience,vol.11211,Springer,2018,pp.833–851

  43. [52]

    UNet++: Nested U-Net for medical image segmentation,

    Z. Zhou et al., “UNet++: Nested U-Net for medical image segmentation,” in Proceedings of DLMIA 2018 and ML-CDS 2018, Granada, Spain, September 20, 2018, Lecture Notes inComputerScience,vol.11045,Springer,2018,pp.3–11

  44. [53]

    Pyramid scene parsing network,

    H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing network,” in Proceedings of CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, IEEE Computer Society,2017,pp.6230–6239

  45. [54]

    Deep high-resolutionrepresentationlearningforvisualrecognition,

    J.Wangetal.,“Deep high-resolutionrepresentationlearningforvisualrecognition,”IEEE Trans.PatternAnal.Mach.Intell.,vol.43,no.10,pp.3349–3364,2021

  46. [55]

    Trseg: Transformer for semantic segmentation,

    Y. Jin, D. Han, and H. Ko, “Trseg: Transformer for semantic segmentation,” Pattern RecognitionLetters,vol.148,pp.29–35,2021

  47. [56]

    Multimodal remote sensing image segmentation with intuition-inspired hypergraph modeling,

    Q. He et al., “Multimodal remote sensing image segmentation with intuition-inspired hypergraph modeling,” IEEE Transactions on Image Processing, vol. 32, pp. 1474–1487, 2023

Pith tools

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