Pith. sign in

REVIEW 3 major objections 4 minor 52 references

Referring Remote Sensing Image Segmentation with Cross-view Semantics Interaction Network

T0 review · 3 major / 4 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 The paper claims that a network processing the full remote-sensing image and zoomed patches in parallel, with stage-by-stage cross-view attention, segments tiny and ambiguous targets better than single-view methods on three benchmarks.

desk verdict The dual-view input idea is real and the evaluation is unusually thorough, but the key ablation never rules out that the gains come from extra pixels and parameters rather than from cross-view interaction. read the letter →

arxiv 2508.01331 v1 pith:OVYOWM27 submitted 2025-08-02 cs.CV cs.AI

Jiaxing Yang , Lihe Zhang , Huchuan Lu This is my paper · ORCID
classification cs.CVcs.AI
keywords referringimagesegmentationremotesensingcross-viewinteractioncross-modalattentionSwinTransformerdilatedsmallobjectRRSIS
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

Referring remote sensing image segmentation asks a model to outline an object in an aerial photo from a sentence describing it. The paper's target problem is that such objects range from whole airports to tiny vehicles, and existing single-view models lose small or ambiguous targets. CSINet answers this by running two parallel streams: a low-resolution full image that supplies global context and high-resolution patches that supply local detail, with a window-attention module passing information between the streams after every encoding stage. The paper claims this design, plus a dilation-based decoder, raises mean IoU on three benchmarks, with the biggest improvements on exactly the small and slender targets that motivated it. If true, the approach gives a simple recipe for making referring segmentation robust to scale variation in overhead imagery.

What carries the argument

The load-bearing objects are CVWin and CDAD. CVWin, the Cross-View Window-attention module, is inserted after every vision encoding stage; it splits the remote-view and close-view features into matching windows and runs two asymmetric cross-attention passes, one feeding close-view detail into the remote branch and one feeding remote-view global semantics into the close branch, then fuses each branch with a feed-forward network. CDAD, the Collaboratively Dilated Attention enhanced Decoder, builds a joint feature from the final stage of both views and lets each branch query it with dilated, row-spaced attention, which is designed to capture long, slim targets oriented in different directions. The stage-by-stage bidirectional exchange is the mechanism the paper credits for unifying global and local semantics.

What would settle it

Re-running the published protocol on the RRSIS-D validation set and comparing CSINet to a single-view baseline that receives the full image plus the same 2x2 close-view patches but no cross-attention would settle the causal claim; if the single-view model matches CSINet's mean IoU, the interaction mechanism is not the source of the gains, and if an independent re-run reproduces the reported test margins, the results are robust.

Watch

Extended reading notes

Core claim

The paper's central discovery is that 'seeing the whole scene' and 'seeing the target up close' are best treated as two parallel feature streams that exchange information at every encoding stage, rather than as scales inside a single stream. The paper supports this with ablations in which removing the close-view branch costs about 12.96 mIoU and removing the remote-view branch costs about 4.05 on RRSIS-D validation, while unidirectional interaction recovers only part of the bidirectional gain. With a Swin-Base backbone, CSINet reports test mIoU of 67.47 on RRSIS-D against 64.20 for RMSIN and 64.24 for CroBIM; 73.88 on RefSegRS against 62.58 for RMSIN; and 69.25 on RIS-Bench against 67.32 for CroBIM. In all three comparisons the mIoU margin is larger than the oIoU margin, which the paper reads as evidence that the method helps most on small targets.

Load-bearing premise

The load-bearing premise is that the internally re-run baselines match their original settings and that the gains come from cross-view interaction rather than from simply feeding extra high-resolution patches.

Editorial extensions

If this is right

  • If the central claim is right, single-view referring segmentation systems can be upgraded by adding a close-view patch stream and stage-wise window attention, without replacing the backbone.
  • The consistently larger mIoU than oIoU gains imply that an evaluation focused only on overall IoU will understate the method's benefit for small objects.
  • The ablations suggest that bidirectional cross-view exchange is better than either direction alone, so future dual-branch designs should keep both information flows.
  • Per-category results on ships, chimneys, vehicles, bridges, and harbors indicate that the design specifically relieves the failure mode named at the outset: tiny, slender, ambiguous targets.

Reading between the lines

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

  • An extension the paper leaves implicit: the same dual-view interaction could apply to other overhead-image tasks, such as building footprint extraction, road mapping, or change detection, where object scale varies sharply.
  • The chosen 2x2 patch split is validated only on RRSIS-D; on higher-resolution or differently shaped imagery a data-dependent patch count would be a natural next experiment.
  • Because the paper does not compare against a single-view model that receives the same high-resolution patches without cross attention, part of the reported gain may come from extra input resolution; isolating that would sharpen the causal story.
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

3 major / 4 minor

Summary. The paper proposes CSINet, a referring remote sensing image segmentation framework that processes a low-resolution full image (remote view) in parallel with high-resolution non-overlapping patches (close views), exchanging information between the two branches through a Cross-View Window-attention module (CVWin) at each encoding stage and decoding with a Collaboratively Dilated Attention enhanced Decoder (CDAD). The method is evaluated on RRSIS-D, RefSegRS, and RIS-Bench, where it reports state-of-the-art results on oIoU, mIoU, and P@X, along with per-category mIoU, ablations of the main components, and speed/parameter comparisons. The paper does not release code or data and reproduces all baselines in-house.

Significance. If the reported gains are reproducible, the multi-view input design is a useful direction for handling scale diversity in remote sensing referring segmentation, and the per-category analysis gives concrete evidence of improvement on small targets. The evaluation is broad: three benchmarks, five metrics, a per-category table, and ablations for each proposed module, plus an explicit limitations section. However, the causal claim that cross-view interaction, rather than the increased input resolution and parameter count, drives the gains is not yet established by the ablation design, and the decoder description is not sufficiently precise to reimplement. These issues are fixable but require additional experiments and exposition.

major comments (3)
  1. [Section IV-D, Table VII] The single-view controls 'Only Remote' and 'Only Close' do not match the full model's input pixel count, parameter count, or FLOPs: 'Only Remote' sees a single 384×384 image, 'Only Close' sees four non-overlapping 384×384 patches without cross-patch interaction, while CSINet sees both the remote image and the four close patches (a total of 768×768 pixels). The reported mIoU gain could therefore be caused by the extra input pixels and the larger model (278.4M vs. 208.9M for RMSIN), rather than by the proposed bidirectional cross-view interaction. To support the central mechanistic claim, please add a matched single-view baseline that processes the full 768×768 image with the same backbone and decoder, and a 'close patches without interaction' baseline with the same total pixel budget and comparable capacity.
  2. [Section III-C, Eqs. (10)-(12)] The CDA formulation is not dimensionally reproducible as written. Specifically, a 1×1 convolution with padding size (H_adjust, 0) and the subsequent slice Kjoint[H_adjust:2H_adjust, :, :, :] in Eq. (11) require a height dimension that is not documented; the stated output size H_adjust × Nslice × (2J+1)Sslice × C does not make the mapping between the height axis, the Nslice split, and the concatenation along width clear. The sentence 'The specific processes are omitted for convenience' further hides a core contribution. Please provide a complete tensor-shape derivation or pseudocode for Eqs. (10)-(12), including how Split, the padding, and the slicing axes interact.
  3. [Section IV-B and Tables I-III] The headline comparisons depend on the authors' in-house reproductions of all baselines, but the paper does not release code or data and gives no per-baseline training protocol (batch size, epochs, learning rate schedule, loss weights, random seeds, number of runs). Because several reported margins are small (for example, +1.93 mIoU over CroBIM on RIS-Bench test), differences in training configuration could change the conclusions. Please release code and trained models, or at minimum provide a detailed reproduction protocol for each baseline and report results across multiple seeds with variance.
minor comments (4)
  1. [Section IV-C, text after Table II] Several claimed gains in the text do not match the tables; for example, the RRSIS-D paragraph reports gains over CARIS of 5.87 (average P@X), 1.37 (oIoU), and 5.05 (mIoU), whereas Table I test values give 6.39, 1.17, and 5.30, and the RefSegRS paragraph reports gains over LGCE of 17.68, 4.09, and 10.3, whereas Table II test values give approximately 18.27, 4.51, and 11.18. Please reconcile all such numbers.
  2. [Eq. (13)] In the decoder recurrence, 'Di = CBR(Cat(Ii, Up(D1)))' appears to be a typo for 'Up(Di-1)' based on the definition of Ii and the removal ablation in Table VIII; please correct the equation.
  3. [Throughout] There are several typos and minor inconsistencies: 'patalleled' in the contributions list, 'parch images' in Section IV-B, 'GroudnTruth' in the Figure 5 caption, and 'Different form' before Eq. (12). In addition, Table IV should state whether the per-category numbers are validation, test, or combined, since the text refers to both sets.
  4. [Figure 5 caption] The caption says 'Quantitative results' but the figure shows segmentation visualizations; this should read 'Qualitative results'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CSINet is a standard supervised learning pipeline trained on one split and evaluated on held-out test splits; no reported prediction reduces to a fitted input or self-citation by construction.

full rationale

The paper's central claim is empirical: CSINet is trained end-to-end on the training splits of RRSIS-D, RefSegRS, and RIS-Bench and evaluated on the official test splits (Tables I–III). The method is not derived from the test metrics, and no parameter is fitted to the test set and then reported as a prediction. Ablations (Tables V–VIII) compare variants on the validation split of RRSIS-D, which is standard practice; selecting hyperparameters such as Nview=2, slice size 5, and J=3 on validation does not make the test-set prediction equivalent to the validation fit. The cross-view interaction modules (CVWin, CDAD) are architectural contributions evaluated by controlled ablations, not by a self-referential derivation. Self-citations to the authors' earlier works appear only in the related-work discussion and as technical inspiration (e.g., the gate mechanism from LAVT [18] and the window concept from Swin [31]); the central architecture and evaluation do not depend on an unverified uniqueness theorem or on a prior result by the same authors that already contains the claimed conclusion. The noted ambiguity in the CDAD equations (Eq. 10–11) is a reproducibility concern, not a circularity concern, because it does not make the output equal to the input by construction. The paper is therefore self-contained against external benchmarks, and no circular step meeting the evidentiary standard is present.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim depends on five free parameters, all chosen by validation or stated without ablation. There are no invented physical entities. The main domain assumptions are that ImageNet/Web-pretrained encoders work for aerial data and that window alignment between views is semantically meaningful. The parameter choices are typical for deep learning papers, but they are still ad hoc choices that affect the reported numbers.

free parameters (5)
  • Nview (number of close-view patches per side) = 2
    Section III-A and Table VI: chosen empirically as the best split pattern on RRSIS-D validation (2x2 beats 3x3 and 4x4). This choice affects the input structure.
  • Slice size Sslice in CDA = 5
    Section IV.B and Table IX: set to 5, giving the best validation oIoU (78.72) among sizes 5-6 tested with J=3..6.
  • Dilation density J in CDA = 3
    Section IV.B: 'the density number J of collaboratively dilation attention is set to 3', inspired by ASPP, and Table IX confirms J=3 with slice 5 is optimal.
  • Decoder channel compression Ccmp = Cvis_4 / 2
    Section IV.B: chosen as half the top-level feature channels; no ablation reported for this choice.
  • Training hyperparameters (epochs, LR, loss weights) = 40/60/40 epochs, LR 5e-5, Dice 0.9/BCE 0.1
    Section IV.B: standard choices stated without sensitivity analysis. Not central to the architecture claim but affect the reported numbers.
assumptions (3)
  • domain assumption Swin Transformer and BERT pretrained weights transfer to remote sensing imagery
    Section III-A: both encoders are initialized from ImageNet-22K and HuggingFace BERT weights, assuming their features are usable for aerial images despite the domain gap.
  • domain assumption The window correspondence assumption: tokens in co-located remote-view and close-view windows refer to the same image content
    Section III-B: this justifies the asymmetrical-window cross-attention. It holds only because the close-view patches tile the remote view exactly after splitting and resizing, and is an architectural premise, not a demonstrated fact.
  • standard math Standard supervised training with Dice+BCE losses optimizes the metrics reported
    Section IV.B: uses standard loss functions for segmentation; this is normal practice and does not require proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Referring Remote Sensing Image Segmentation with Cross-view Semantics Interaction Network." pith.science (2026). https://pith.science/paper/OVYOWM27

@misc{pith2026250801331,
  author       = {Pith},
  title        = {Pith review of: Referring Remote Sensing Image Segmentation with Cross-view Semantics Interaction Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OVYOWM27}},
  note         = {Machine review of arXiv:2508.01331}
}
read the original abstract

Recently, Referring Remote Sensing Image Segmentation (RRSIS) has aroused wide attention. To handle drastic scale variation of remote targets, existing methods only use the full image as input and nest the saliency-preferring techniques of cross-scale information interaction into traditional single-view structure. Although effective for visually salient targets, they still struggle in handling tiny, ambiguous ones in lots of real scenarios. In this work, we instead propose a paralleled yet unified segmentation framework Cross-view Semantics Interaction Network (CSINet) to solve the limitations. Motivated by human behavior in observing targets of interest, the network orchestrates visual cues from remote and close distances to conduct synergistic prediction. In its every encoding stage, a Cross-View Window-attention module (CVWin) is utilized to supplement global and local semantics into close-view and remote-view branch features, finally promoting the unified representation of feature in every encoding stage. In addition, we develop a Collaboratively Dilated Attention enhanced Decoder (CDAD) to mine the orientation property of target and meanwhile integrate cross-view multiscale features. The proposed network seamlessly enhances the exploitation of global and local semantics, achieving significant improvements over others while maintaining satisfactory speed.

Figures

Figures reproduced from arXiv: 2508.01331 by the authors.

Figure 1
Figure 1. Architectures for referring remote sensing image seg [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture overview of CSINet. It feeds on a complete image that simulates remote view and several image patches [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The details of CDA. The upper part shows how remote [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization results of proposed CSINet and other state of the art algorithms from the validation set of RRSIS-D. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Quantitative results. The symbols Only R2C and Only [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 50 canonical work pages

  1. [1]

    Segmentation from natural language expressions,

    R. Hu, M. Rohrbach, and T. Darrell, “Segmentation from natural language expressions,” in Proc. Eur . Conf. Comput. Vis. Springer, 2016, pp. 108–124

    work page 2016

  2. [2]

    Referring image segmentation via recurrent refinement networks,

    R. Li, K. Li, Y .-C. Kuo, M. Shu, X. Qi, X. Shen, and J. Jia, “Referring image segmentation via recurrent refinement networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2018, pp. 5745–5753

    work page 2018

  3. [3]

    Recurrent multimodal interaction for referring image segmentation,

    C. Liu, Z. Lin, X. Shen, J. Yang, X. Lu, and A. Yuille, “Recurrent multimodal interaction for referring image segmentation,” in Proc. IEEE Int. Conf. Comput. Vis. , 2017, pp. 1271–1280

    work page 2017

  4. [4]

    Mdetr-modulated detection for end-to-end multi-modal understanding,

    A. Kamath, M. Singh, Y . LeCun, G. Synnaeve, I. Misra, and N. Carion, “Mdetr-modulated detection for end-to-end multi-modal understanding,” in Proc. IEEE Int. Conf. Comput. Vis. , 2021, pp. 1780–1790

    work page 2021

  5. [5]

    Key-word-aware network for referring expression image segmentation,

    H. Shi, H. Li, F. Meng, and Q. Wu, “Key-word-aware network for referring expression image segmentation,” in Proc. Eur . Conf. Comput. Vis., 2018, pp. 38–54

    work page 2018

  6. [6]

    Cross-modal self-attention network for referring image segmentation,

    L. Ye, M. Rochan, Z. Liu, and Y . Wang, “Cross-modal self-attention network for referring image segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2019, pp. 10 502–10 511

    work page 2019

  7. [7]

    Linguistic structure guided context modeling for referring image segmentation,

    T. Hui, S. Liu, S. Huang, G. Li, S. Yu, F. Zhang, and J. Han, “Linguistic structure guided context modeling for referring image segmentation,” in Proc. Eur . Conf. Comput. Vis. Springer, 2020, pp. 59–75

    work page 2020

  8. [8]

    Vision-language transformer and query generation for referring segmentation,

    H. Ding, C. Liu, S. Wang, and X. Jiang, “Vision-language transformer and query generation for referring segmentation,” in Proc. IEEE Int. Conf. Comput. Vis. , 2021, pp. 16 321–16 330

    work page 2021

Show all 52 references
  1. [9]

    Bi-directional relationship inferring network for referring image segmentation,

    Z. Hu, G. Feng, J. Sun, L. Zhang, and H. Lu, “Bi-directional relationship inferring network for referring image segmentation,” inProc. IEEE Conf. Comput. Vis. Pattern Recog. , 2020, pp. 4424–4433

    2020

  2. [10]

    Mattnet: Modular attention network for referring expression compre- hension,

    L. Yu, Z. Lin, X. Shen, J. Yang, X. Lu, M. Bansal, and T. L. Berg, “Mattnet: Modular attention network for referring expression compre- hension,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2018, pp. 1307–1315

    2018

  3. [11]

    Multi-task collaborative network for joint referring expression comprehension and segmentation,

    G. Luo, Y . Zhou, X. Sun, L. Cao, C. Wu, C. Deng, and R. Ji, “Multi-task collaborative network for joint referring expression comprehension and segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2020, pp. 10 034–10 043

    2020

  4. [12]

    Locate then segment: A strong pipeline for referring image segmentation,

    Y . Jing, T. Kong, W. Wang, L. Wang, L. Li, and T. Tan, “Locate then segment: A strong pipeline for referring image segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2021, pp. 9858–9867

    2021

  5. [13]

    Instance-specific feature propagation for referring segmentation,

    C. Liu, X. Jiang, and H. Ding, “Instance-specific feature propagation for referring segmentation,” IEEE Trans Multimedia , 2022

    2022

  6. [14]

    Polyformer: Referring image segmentation as sequential polygon generation,

    J. Liu, H. Ding, Z. Cai, Y . Zhang, R. K. Satzoda, V . Mahadevan, and R. Manmatha, “Polyformer: Referring image segmentation as sequential polygon generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2023, pp. 18 653–18 663

    2023

  7. [15]

    Restr: Convolution- free referring image segmentation using transformers,

    N. H. Kim, D. Kim, C. Lan, W. Zeng, and S. Kwak, “Restr: Convolution- free referring image segmentation using transformers,” Proc. IEEE Conf. Comput. Vis. Pattern Recog. , pp. 18 124–18 133, 2022

    2022

  8. [16]

    Bilateral knowledge interaction network for referring image segmentation,

    H. Ding, S. Zhang, Q. Wu, S. Yu, J. Hu, L. Cao, and R. Ji, “Bilateral knowledge interaction network for referring image segmentation,” IEEE Trans Multimedia, 2023

    2023

  9. [17]

    Encoder fusion network with co- attention embedding for referring image segmentation,

    G. Feng, Z. Hu, L. Zhang, and H. Lu, “Encoder fusion network with co- attention embedding for referring image segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2021, pp. 15 506–15 515

    2021

  10. [18]

    Lavt: Language-aware vision transformer for referring image segmentation,

    Z. Yang, J. Wang, Y . Tang, K. Chen, H. Zhao, and P. H. Torr, “Lavt: Language-aware vision transformer for referring image segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2022, pp. 18 155– 18 165

    2022

  11. [19]

    Semantics-aware dynamic localization and refinement for referring image segmentation,

    Z. Yang, J. Wang, Y . Tang, K. Chen, H. Zhao, and H. Torr, Philip, “Semantics-aware dynamic localization and refinement for referring image segmentation,” in Proc. AAAI Conf. Artif. Intell. , 2023

    2023

  12. [20]

    Referring image segmentation with fine- grained semantic funneling infusion,

    J. Yang, L. Zhang, and H. Lu, “Referring image segmentation with fine- grained semantic funneling infusion,” IEEE Trans. Neural Netw. Learn. Syst. IEEE Trans. Neural Networks , 2023

    2023

  13. [21]

    Contrastive grouping with transformer for referring image segmentation,

    J. Tang, G. Zheng, C. Shi, and S. Yang, “Contrastive grouping with transformer for referring image segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2023, pp. 23 570–23 580

    2023

  14. [22]

    Towards robust referring image segmentation,

    J. Wu, X. Li, X. Li, H. Ding, Y . Tong, and D. Tao, “Towards robust referring image segmentation,” IEEE Trans. Image Process. , 2024

    2024

  15. [23]

    Fully and weakly supervised referring expression segmentation with end-to-end learning,

    H. Li, M. Sun, J. Xiao, E. G. Lim, and Y . Zhao, “Fully and weakly supervised referring expression segmentation with end-to-end learning,” IEEE Trans. Circuits Syst. Video Technol. , vol. 33, no. 10, pp. 5999– 6012, 2023

    2023

  16. [24]

    Cross-modal recurrent semantic comprehension for referring image segmentation,

    C. Shang, H. Li, H. Qiu, Q. Wu, F. Meng, T. Zhao, and K. N. Ngan, “Cross-modal recurrent semantic comprehension for referring image segmentation,” IEEE Trans. Circuits Syst. Video Technol. , vol. 33, no. 7, pp. 3229–3242, 2022

    2022

  17. [25]

    Cmirnet: Cross- modal interactive reasoning network for referring image segmentation,

    M. Xu, T. Xiao, Y . Liu, H. Tang, Y . Hu, and L. Nie, “Cmirnet: Cross- modal interactive reasoning network for referring image segmentation,” IEEE Trans. Circuits Syst. Video Technol., vol. 35, no. 4, pp. 3234–3249, 2025

    2025

  18. [26]

    Exploring fine-grained image-text alignment for referring remote sensing image segmentation,

    S. Lei, X. Xiao, T. Zhang, H.-C. Li, Z. Shi, and Q. Zhu, “Exploring fine-grained image-text alignment for referring remote sensing image segmentation,” IEEE Trans. Geosci. Remote Sens. , 2024

    2024

  19. [27]

    Rrsis: Referring remote sensing image segmentation,

    Z. Yuan, L. Mou, Y . Hua, and X. X. Zhu, “Rrsis: Referring remote sensing image segmentation,” IEEE Trans. Geosci. Remote Sens. , 2024

    2024

  20. [28]

    Rethinking the implicit optimization paradigm with dual alignments for referring remote sensing image segmentation,

    Y . Pan, R. Sun, Y . Wang, T. Zhang, and Y . Zhang, “Rethinking the implicit optimization paradigm with dual alignments for referring remote sensing image segmentation,” in Proc. ACM Int. Conf. Multimedia. , 2024, pp. 2031–2040

    2024

  21. [29]

    Rotated multi-scale interaction network for referring remote sensing image segmentation,

    S. Liu, Y . Ma, X. Zhang, H. Wang, J. Ji, X. Sun, and R. Ji, “Rotated multi-scale interaction network for referring remote sensing image segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2024, pp. 26 658–26 668

    2024

  22. [30]

    Cross-modal bidirectional inter- action model for referring remote sensing image segmentation,

    Z. Dong, Y . Sun, Y . Gu, and T. Liu, “Cross-modal bidirectional inter- action model for referring remote sensing image segmentation,” arXiv preprint arXiv:2410.08613, 2024

    2024 arXiv

  23. [31]

    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 Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2021, pp. 10 012–10 022

    2021

  24. [32]

    Referring image seg- mentation by generative adversarial learning,

    S. Qiu, Y . Zhao, J. Jiao, Y . Wei, and S. Wei, “Referring image seg- mentation by generative adversarial learning,” IEEE Trans Multimedia , vol. 22, no. 5, pp. 1333–1344, 2019

    2019

  25. [33]

    Lisa: Reasoning segmentation via large language model,

    X. Lai, Z. Tian, Y . Chen, Y . Li, Y . Yuan, S. Liu, and J. Jia, “Lisa: Reasoning segmentation via large language model,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2024, pp. 9579–9589

    2024

  26. [34]

    Fine-grained video-text retrieval with hierarchical graph reasoning,

    S. Chen, Y . Zhao, Q. Jin, and Q. Wu, “Fine-grained video-text retrieval with hierarchical graph reasoning,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2020, pp. 10 638–10 647

    2020

  27. [35]

    Learning dual semantic relations with graph attention for image-text matching,

    K. Wen, X. Gu, and Q. Cheng, “Learning dual semantic relations with graph attention for image-text matching,” IEEE Trans. Circuits Syst. Video Technol., vol. 31, no. 7, pp. 2866–2879, 2020

    2020

  28. [36]

    Show, attend and tell: Neural image caption generation with visual attention,

    K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhudinov, R. Zemel, and Y . Bengio, “Show, attend and tell: Neural image caption generation with visual attention,” in Proc. Int. Conf. Mach. Learn. PMLR, 2015, pp. 2048–2057

    2015

  29. [37]

    Multimodal transformer with multi- view visual representation for image captioning,

    J. Yu, J. Li, Z. Yu, and Q. Huang, “Multimodal transformer with multi- view visual representation for image captioning,” IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 12, pp. 4467–4480, 2019

    2019

  30. [38]

    Transformer- based language-person search with multiple region slicing,

    H. Li, J. Xiao, M. Sun, E. G. Lim, and Y . Zhao, “Transformer- based language-person search with multiple region slicing,” IEEE Trans. Circuits Syst. Video Technol. , vol. 32, no. 3, pp. 1624–1633, 2021

    2021

  31. [39]

    Language-guided navigation via cross-modal grounding and alternate adversarial learning,

    W. Zhang, C. Ma, Q. Wu, and X. Yang, “Language-guided navigation via cross-modal grounding and alternate adversarial learning,” IEEE Trans. Circuits Syst. Video Technol. , vol. 31, no. 9, pp. 3469–3481, 2020

    2020

  32. [40]

    Mask grounding for referring image segmentation,

    Y . X. Chng, H. Zheng, Y . Han, X. Qiu, and G. Huang, “Mask grounding for referring image segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2024

    2024

  33. [41]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” in International Conference on Learning Representations , 2022. [Online]. Available: https://openreview.net/forum?id=nZeVKeeFYf9

    2022

  34. [42]

    Visual grounding in remote sensing images,

    Y . Sun, S. Feng, X. Li, Y . Ye, J. Kang, and X. Huang, “Visual grounding in remote sensing images,” inProceedings of the 30th ACM International Conference on Multimedia , 2022, pp. 404–412

    2022

  35. [43]

    Rsvg: Exploring data and models for visual grounding on remote sensing data,

    Y . Zhan, Z. Xiong, and Y . Yuan, “Rsvg: Exploring data and models for visual grounding on remote sensing data,” IEEE Trans. Geosci. Remote Sens., vol. 61, pp. 1–13, 2023

    2023

  36. [44]

    Language-guided progressive attention for visual grounding in remote sensing images,

    K. Li, D. Wang, H. Xu, H. Zhong, and C. Wang, “Language-guided progressive attention for visual grounding in remote sensing images,” IEEE Trans. Geosci. Remote Sens. , vol. 62, pp. 1–13, 2024

    2024

  37. [45]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  38. [46]

    Refer- ring image segmentation via cross-modal progressive comprehension,

    S. Huang, T. Hui, S. Liu, G. Li, Y . Wei, J. Han, L. Liu, and B. Li, “Refer- ring image segmentation via cross-modal progressive comprehension,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog. , 2020, pp. 10 488– 10 497. 12

    2020

  39. [47]

    Cross-modal progressive comprehension for referring segmentation,

    S. Liu, T. Hui, S. Huang, Y . Wei, B. Li, and G. Li, “Cross-modal progressive comprehension for referring segmentation,” IEEE Trans. Pattern Anal. Mach. Intell. , 2021

    2021

  40. [48]

    Cris: Clip-driven referring image segmentation,

    Z. Wang, Y . Lu, Q. Li, X. Tao, Y . Guo, M. Gong, and T. Liu, “Cris: Clip-driven referring image segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 11 686–11 695

    2022

  41. [49]

    Caris: Context-aware referring image segmentation,

    S.-A. Liu, Y . Zhang, Z. Qiu, H. Xie, Y . Zhang, and T. Yao, “Caris: Context-aware referring image segmentation,” in Proc. ACM Int. Conf. Multimedia., 2023

    2023

  42. [50]

    Beyond one-to-one: Rethinking the referring image segmentation,

    Y . Hu, Q. Wang, W. Shao, E. Xie, Z. Li, J. Han, and P. Luo, “Beyond one-to-one: Rethinking the referring image segmentation,” in Proc. IEEE Int. Conf. Comput. Vis. , October 2023, pp. 4067–4077

    2023

  43. [51]

    Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation,

    Z. Xu, Z. Chen, Y . Zhang, Y . Song, X. Wan, and G. Li, “Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation,” in Proc. IEEE Int. Conf. Comput. Vis. , 2023, pp. 17 503–17 512

    2023

  44. [52]

    Spatial pyramid pooling in deep convolutional networks for visual recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Spatial pyramid pooling in deep convolutional networks for visual recognition,”IEEE Trans. Pattern Anal. Mach. Intell. , vol. 37, no. 9, pp. 1904–1916, 2015

    1904

Pith tools

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