Pith. sign in

REVIEW 3 major objections 6 minor 80 references

CROSS: Cascaded Distillation and Dual-Constraint Grounding for Remote Sensing Referring Segmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read CROSS claims that remote sensing referring segmentation improves when SAM's patch-affinity structure is fed into the VLM's intermediate layers as a text-gated soft regularizer and the model is contrastively trained against masked…

desk verdict Solid empirical RRSIS paper with real gains at high IoU thresholds, but the 'text-guided filtering' mechanism in LGCD is mathematically overstated: the loss reweights all pairs rather than selecting linguistically relevant topology. read the letter →

arxiv 2608.03147 v2 pith:WPU336TI submitted 2026-08-04 cs.CV

Tingzhang Luo , Ruizhong Liu , Yichao Liu , Cheng Fan , Yu Liu , Jianyuan Guo This is my paper · ORCID
classification cs.CV
keywords referringremotesensingimagesegmentationvision-languagemodelSegmentAnythingknowledgedistillationcontrastivelearningspatialreasoningGrammatrix
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

CROSS claims that existing pipelines that pair a vision-language model with the Segment Anything Model (SAM) for referring remote sensing segmentation fail for two structural reasons: the VLM prompts SAM in only one direction, so SAM's pixel-level geometry never guides the prompt, and the VLM's contrastive pretraining makes it lean on object category while ignoring spatial modifiers. To fix this, CROSS distills SAM's patch-to-patch affinity structure into three intermediate layers of SigLIP 2, using a text-derived soft mask so only linguistically relevant geometry is transferred, and adds a contrastive loss that mines background distractors and synthetically inverted spatial phrases. The reported results are state of the art on both standard benchmarks, with cIoU of 83.25 on RefSegRS and 79.89 on RRSIS-D, and the largest margins over the previous best method appear at the strictest overlap threshold, Pr@0.9. If the claim holds, it means structural priors can be injected into a VLM without using SAM as a mask oracle, and spatial reasoning can be trained directly rather than left to chance.

What carries the argument

Linguistic-Guided Cascaded Distillation (LGCD) is the main mechanism. At three chosen SigLIP 2 blocks, the visual features are refined by cross-attention with text, and the cross-attention matrix is averaged over tokens to form a soft mask M_text(p). The distillation loss compares Gram matrices G(H_i) and G(S), where the Gram matrix is the matrix of pairwise cosine similarities between spatial patch features and S is SAM 2's terminal encoder feature; the text mask weights each patch pair, so only linguistically relevant topology is aligned. Perspective-Spatial Contrastive Learning (PSCL) is the complementary mechanism: the ground-truth mask blinds the true object, the top-K background patches most similar to the text become hard negatives, and rule-based or LLM-generated counterfactual texts with inverted spatial relations provide a second negative term in an asymmetric contrastive loss. The two losses are added to the mask losses with weights 0.5 and 0.2, and the cascaded layers are set to indices 9, 18, and 27.

What would settle it

Take the exact CROSS training recipe and replace SAM's terminal feature with a fixed random or patch-shuffled feature map while keeping the text mask and all losses; if cIoU on RefSegRS still rises by the reported margin, the distillation is not carrying structural information. Alternatively, build a perturbation test set that swaps spatial words in the referring expressions and measure mask IoU with and without PSCL; if the robustness gap over the baseline disappears, the contrastive loss is not the source of spatial grounding.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the two bottlenecks—architectural weak-coupling and object-centric semantic bias—can be addressed in one framework without changing the frozen SigLIP 2 and SAM 2 backbones. LGCD aligns the Gram matrices of VLM intermediate features with the Gram matrix of SAM 2's encoder output, weighted by a cross-attention text mask, which concentrates the dense prompt and yields the largest cIoU gains; PSCL then shatters the object-category shortcut by pushing away mask-filtered background patches that most resemble the text and by penalizing the positive visual prototype against spatially inverted counterfactual texts. With both modules, the paper reports cIoU gains from 80.82 to 83.25 on RefSegRS and from 77.80 to 79.89 on RRSIS-D over the SigLIP 2 + SAM 2 baseline, and the ablation shows each component contributes, with PSCL adding a smaller but consistent increment. The paper also claims the resulting model remains logically sound under spatial perturbations such as swapping left and right, where the baseline collapses to category-level search.

Load-bearing premise

The load-bearing premise is that SAM's patch-to-patch affinity geometry and SigLIP 2's intermediate features are comparable enough that aligning their Gram matrices under a text-derived mask improves rather than corrupts the VLM's semantics, and the paper offers no direct measurement of that cross-encoder alignment.

Editorial extensions

If this is right

  • An explicit text-gated relational loss can transfer geometry from SAM into a VLM without copying masks as hard labels.
  • The method keeps the training-time distillation off the inference path, so the reported spatial-concentration gains do not cost inference speed.
  • The largest accuracy gains occur at Pr@0.9, so the practical consequence is cleaner boundaries and fewer over- or under-segmentations in cluttered scenes.
  • Only 7.2 percent of total parameters are trainable, which means the recipe is compatible with parameter-efficient fine-tuning of billion-scale backbones.
  • Spatial counterfactual contrast is stable across different LLM generators, making the approach reproducible without a particular proprietary model.

Reading between the lines

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

  • Nothing in the LGCD mechanism is specific to remote sensing; the same text-gated Gram distillation could be applied to any task where a language-image encoder is coupled to a promptable mask decoder, such as natural-image referring segmentation or open-vocabulary segmentation.
  • The effectiveness of LGCD rests on the cross-attention mask pointing at the referred object; if the mask is diffuse early in training, the distillation could reinforce background structure instead of suppressing it, so monitoring mask sharpness during training would be a natural diagnostic.
  • The spatial counterfactual rules cover absolute positions and subject-object swaps; extending them to temporal relations such as before or after, or size relations such as larger or smaller, would test whether the same contrastive logic generalizes beyond the current rule set.
  • A cleaner test of whether relational topology is the active ingredient would be to ablate the Gram alignment against a masked pixel-wise MSE on the same soft-mask regions; the paper does not report that comparison.
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 / 6 minor

Summary. The paper proposes CROSS, a training-time framework for remote sensing referring image segmentation built on SigLIP 2 and SAM 2. It targets two stated bottlenecks: architectural weak coupling between the VLM and the SAM decoder, and object-centric semantic bias inherited from VLM pretraining. The first contribution, LGCD, extracts intermediate visual features from three SigLIP stages, refines them with cross-attention to text, and distills SAM 2 Gram-matrix affinities into those features using a per-pair text-derived soft weight. The second, PSCL, adds an InfoNCE-style loss with top-K background distractors and synthetically generated spatial-counterfactual texts as negatives. Experiments on RefSegRS and RRSIS-D report cIoU 83.25 and 79.89, respectively, with larger gains at Pr@0.9 over RSRefSeg 2, together with ablations, hyperparameter sensitivity analyses, parameter counts, and qualitative robustness visualizations.

Significance. If the mechanism claims are taken as stated, the paper would be a useful contribution: it demonstrates parameter-efficient tuning (7.2% trainable parameters), provides comprehensive ablations of components, layer choices, negative types, and LLM generators, and includes additional training runs in the appendix. The reported gains at high-IoU thresholds are practically relevant. However, the significance is conditional. The central novelty claim, that LGCD performs text-guided selection of SAM's topology, is not supported by the loss formula in Eq. (6), which applies a bounded positive weight to every pair rather than filtering out background pairs. Similarly, the robustness-to-perturbation claim is supported mainly by qualitative examples and by training-time counterfactuals from the same rule family used in the demonstration. With additional targeted experiments and moderated claims, the paper could become a solid journal contribution.

major comments (3)
  1. [Sec. 4.1, Eq. (6), Table 3(a)] The claim that LGCD performs 'text-guided filtering' or 'conditionally routes only linguistically relevant topology' is not supported by the loss definition. The soft weight w^i_{p,q} = alpha + (1-alpha)*(M_text(p)+M_text(q))/2 lies in [0.1, 1] for every pair, so every pair's squared Gram error contributes to the loss; background-involving pairs are down-weighted by at most a factor of 10, not excluded. The loss therefore re-scales the full Gram difference with a spatially smooth positive mask, rather than selecting a subset of pairs. Consequently, the +0.60 cIoU gain attributed to TGD in Table 3(a) is not evidence for a text-selective topology transfer mechanism. Please either add an ablation with unweighted distillation, a random mask, or a binary top-x% mask, or revise the mechanism description to match what Eq. (6) actually computes.
  2. [Sec. 5.5 and Sec. 4.2] The abstract and conclusion claim that CROSS 'maintains precise localization even under severe spatial description perturbations,' but the paper provides no quantitative perturbation benchmark. Section 5.5 gives only two qualitative examples (Fig. 4), and the training-time PSCL counterfactuals of Sec. 4.2 are generated with the same rule family (absolute inversion, subject-object swapping) used to demonstrate robustness, making the demonstration partially self-referential. Please add a quantitative evaluation on a held-out perturbation set, for example reporting cIoU and Pr@0.5/0.9 before and after spatial-text perturbations for CROSS and RSRefSeg 2, or temper the robustness claim to what the current evidence supports.
  3. [Tables 1-2 and Appendix Table 8] The central state-of-the-art claim is based on single-run numbers. The appendix reports three runs for CROSS on RRSIS-D but gives no variance for the strongest baseline, RSRefSeg 2, so the reported advantages at Pr@0.5 (+1.01) and cIoU (+0.44) may be within run-to-run noise. In addition, Table 2 shows that CROSS does not beat RSRefSeg 2 on all metrics: gIoU is 68.92 versus 69.17 and Pr@0.6 is 74.56 versus 75.78. The text's 'state-of-the-art performance on RRSIS-D' needs qualification to the metrics on which the method actually leads. Please report mean and standard deviation over multiple runs for CROSS and the principal baseline, and state explicitly which metrics are primary for the SOTA claim.
minor comments (6)
  1. [Eq. (7) vs Eq. (6)] The index notation is inconsistent: Eq. (6) defines L^i_distill while Eq. (7) sums L^(k)_distill; please unify the superscript conventions.
  2. [Table 2] The row label 'RSRefSeg-2 [3]' appears to cite the wrong reference; it should be [2] to match RSRefSeg 2, and the corresponding entry in the main text should be checked for consistency.
  3. [Sec. 5.3] The phrase 'Filter-Refine-Verify paradigm' is introduced in the RRSIS-D discussion but is not defined or used elsewhere; either define this terminology in the methodology or remove it.
  4. [Sec. 5.5, Fig. 6] The CKA analysis is presented only as heatmaps with qualitative descriptions of 'stable alignment' and 'smoother transitions'; reporting quantitative layer-pair CKA values would make the claim reproducible and easier to assess.
  5. [Sec. 5.4] The text says a larger alpha (e.g., 0.5) leads to 'excessive suppression of potential target features,' while Sec. 4.1 says alpha maintains background consistency; please clarify the relationship between alpha, background weighting, and target suppression.
  6. [Appendix Table 8] The additional training runs are useful, but they are reported for CROSS alone; including the corresponding runs for the strongest baseline would substantially strengthen the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SOTA is measured on held-out test splits; auxiliary losses are independent empirical regularizers.

full rationale

CROSS does not contain a derivation that reduces to its own inputs. Eq. 6 defines an auxiliary distillation loss comparing the VLM's patch-pair Gram matrix to SAM's; Eq. 8 defines a contrastive loss with mined negatives; Eq. 9 combines them with segmentation losses. None of these quantities is simultaneously the input and the output of the claimed prediction. The headline cIoU/gIoU/Pr numbers are computed on test splits of RefSegRS and RRSIS-D, so the SOTA claim is externally benchmarked rather than constructed. The spatial-perturbation robustness claim is a generalization claim supported by PSCL's counterfactual training; although no independent perturbation benchmark is supplied, the paper does not report a score on the same synthetic counterfactuals used for training as if it were a held-out prediction. The skeptical concern that w in Eq. 6 rescales all pair errors rather than hard-filtering background pairs is a mechanistic validity question, not a circularity: the loss still aligns VLM to SAM and its effect is measured by ablation on held-out cIoU. Self-citations (e.g., [35], [40]-[42], [68]) appear only in related-work and context, and no load-bearing premise rests on an unverified self-citation or uniqueness claim.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The central results rest on tuned hyperparameters and on assumptions about the transferability of foundation-model features, plus assumptions about dataset fidelity. No theoretical constants are derived, so the free-parameter count reflects experimental choices rather than fitted physical quantities.

free parameters (10)
  • lambda1 (distillation loss weight) = 0.5
    Chosen by sensitivity analysis on RefSegRS; larger values destabilize joint optimization.
  • lambda2 (PSCL loss weight) = 0.2
    Tuned via sensitivity analysis on RefSegRS.
  • alpha (soft mask floor) = 0.1
    Grid-searched; larger values suppress potential target features.
  • K (top-k visual negatives) = 8
    Tuned via sensitivity analysis.
  • eta (spatial penalty scaling) = 2
    Ablated in Appendix Fig. 10; values 1, 5, and 10 are worse.
  • layer indices for distillation = {9,18,27}
    Selected by ablation in Table 4; authors say not strictly optimal.
  • LoRA rank r = 16
    Follows the PEFT setting of RSRefSeg 2, not independently tuned.
  • temperature tau = 0.07
    Taken from prior contrastive learning practice.
  • lambda_ce and lambda_dice = 5.0
    Follows RSRefSeg 2 weighting, not separately optimized.
  • training schedule = lr 1e-4, batch 8, 300 epochs
    Fixed settings from implementation; no schedule ablation is reported.
assumptions (6)
  • domain assumption Pre-trained SigLIP 2 and SAM 2 retain useful features for remote sensing after LoRA tuning.
    The pipeline keeps most backbone weights frozen and assumes the foundation models transfer to satellite imagery.
  • domain assumption SAM encoder's Gram affinity matrix encodes spatial topology useful for localization.
    LGCD's distillation loss (Eq. 6) assumes pairwise affinity alignment transfers structural priors; no direct validation is provided.
  • domain assumption The text cross-attention mask in Eq. 4 identifies linguistically relevant spatial regions.
    M_text routes the distillation by averaging attention over all text tokens; attention quality is not evaluated.
  • domain assumption LLM-generated spatial counterfactuals change only spatial relations, not object identity.
    Noun swapping in Rule 2a could alter the entity; if it does, PSCL would penalize correct semantic matches.
  • domain assumption Object-centric bias is the dominant failure mechanism in current VLM-SAM pipelines.
    The paper's diagnostic experiments are qualitative and do not isolate alternate causes.
  • domain assumption RefSegRS and RRSIS-D splits and labels are accurate and compatible with published comparison numbers.
    No annotation audit or data leak check is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CROSS: Cascaded Distillation and Dual-Constraint Grounding for Remote Sensing Referring Segmentation." pith.science (2026). https://pith.science/paper/WPU336TI

@misc{pith2026260803147,
  author       = {Pith},
  title        = {Pith review of: CROSS: Cascaded Distillation and Dual-Constraint Grounding for Remote Sensing Referring Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WPU336TI}},
  note         = {Machine review of arXiv:2608.03147}
}
read the original abstract

Referring Remote Sensing Image Segmentation (RRSIS) has achieved significant progress through the integration of VLMs and the Segment Anything Model (SAM). However, this progress largely relies on strong pre-trained capabilities, while leaving two fundamental limitations insufficiently addressed: (1) Architectural Weak-Coupling, where the unidirectional flow forces reliance on coarse VLM prompts and wastes SAM's pixel-level structural guidance, causing localization drift; and (2) Object-Centric Semantic Bias, where models overemphasize dominant object semantics while remaining insensitive to spatial reasoning crucial for RRSIS. Motivated by these observations, we propose CROSS, a tightly integrated paradigm for RRSIS. First, we introduce Linguistic-Guided Cascaded Distillation (LGCD) to bridge the architectural gap, which distills SAM's geometric affinities as soft regularizers into VLM intermediate layers, injecting dense structural priors to refine localization. Second, Perspective-Spatial Contrastive Learning (PSCL) imposes cross-anchored constraints by mining mask-filtered deceptive distractors and spatial-linguistic counterfactuals as hard negatives, explicitly shattering semantic shortcuts to enforce genuine logical consistency. Extensive experiments on RRSIS benchmarks demonstrate that CROSS achieves state-of-the-art performance and maintains precise localization even under severe spatial description perturbations, standing as a robust new paradigm for RRSIS.

Figures

Figures reproduced from arXiv: 2608.03147 by the authors.

Figure 1
Figure 1. Motivation of the CROSS framework. (Left) Architectural Paradigm Comparison: (a) Previous weakly-coupled pipelines treat VLM and SAM as iso￾lated, fragmented modules, where SAM merely serves as a passive executor of ex￾plicit prompts. (b) Our deeply-coupled CROSS performs text-guided distillation of SAM-derived spatial affinity matrices into the VLM to enforce structural constraints. (Right) Spatial Logic Probing: (… view at source ↗
Figure 2
Figure 2. Overview of the proposed CROSS. Our architecture integrates Linguistic￾Guided Cascaded Distillation (LGCD) to inject SAM’s structural priors into hierarchi￾cal VLM layers, and Perspective-Spatial Contrastive Learning (PSCL) which constructs visual and spatial negative samples to enhance spatial-semantic sensitivity. between the visual feature map and the textual embeddings. This prompt P serves as a spatial prior an… view at source ↗
Figure 3
Figure 3. Sensitivity analysis on RefSegRS with respect to cIoU. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visualization result on RRSIS-D. The targets and spatial descriptions are high￾lighted in red and gray, respectively. Compared to RSRefSeg 2, our method achieves more accurate spatial referring and precise boundaries. By penalizing "logical collapse," CROSS effectively…
Figure 5
Figure 5. Figure 5: Visualization of dense prompt heatmaps. Compared to the baseline RSRefSeg 2, the heatmaps of CROSS are significantly more concen￾trated [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 7
Figure 7. Figure 7: Cases of spatial counterfactual sample construction. 1.2 Spatial Counterfactual Cases [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Additional Qualitative Visualizations of Dense Prompts. [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]
Figure 9
Figure 9. Figure 9: Qualitative Segmentation Results on the RefSegRS Dataset. [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]
Figure 10
Figure 10. Figure 10: Sensitivity Analysis of the Contrastive Hyperparameter [PITH_FULL_IMAGE:figures/full_fig_p027_10.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

80 extracted references · 2 canonical work pages

  1. [3]

    In: IGARSS 2025-2025 IEEE Inter- national Geoscience and Remote Sensing Symposium

    Chen, K., Zhang, J., Liu, C., Zou, Z., Shi, Z.: Rsrefseg: Referring remote sensing image segmentation with foundation models. In: IGARSS 2025-2025 IEEE Inter- national Geoscience and Remote Sensing Symposium. pp. 1070–1074. IEEE (2025)

    2025

  2. [2]

    IEEE Trans- actions on Geoscience and Remote Sensing64, 1–20 (2026).https://doi.org/10

    Chen, K., Liu, C., Chen, B., Zhang, J., Zou, Z., Shi, Z.: Rsrefseg 2: Decoupling referring remote sensing image segmentation with foundation models. IEEE Trans- actions on Geoscience and Remote Sensing64, 1–20 (2026).https://doi.org/10. 1109/TGRS.2025.3647535

    work page arXiv 2026

  3. [1]

    arXiv preprint arXiv:2511.21631 (2025) 16 T

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025) 16 T. Luo et al

    arXiv 2025

  4. [4]

    arXiv preprint arXiv:2305.06558 (2023)

    Cheng, Y., Li, L., Xu, Y., Li, X., Yang, Z., Wang, W., Yang, Y.: Segment and track anything. arXiv preprint arXiv:2305.06558 (2023)

    arXiv 2023

  5. [5]

    Advances in Neural Information Processing Systems 38, 23475–23537 (2026)

    Cho, J.H., Madotto, A., Mavroudi, E., Afouras, T., Nagarajan, T., Maaz, M., Song, Y., Ma, T., Hu, S., Jain, S., et al.: Perceptionlm: Open-access data and models for detailed visual understanding. Advances in Neural Information Processing Systems 38, 23475–23537 (2026)

    2026

  6. [6]

    IEEE Trans- actions on Multimedia26, 5823–5833 (2023)

    Cho, Y., Yu, H., Kang, S.J.: Cross-aware early fusion with stage-divided vision and language transformer encoders for referring image segmentation. IEEE Trans- actions on Multimedia26, 5823–5833 (2023)

    2023

  7. [7]

    IEEE Transactions on Geoscience and Remote Sensing (2025)

    Gao, J., Zhang, D., Wang, F., Ning, L., Zhao, Z., Li, X.: Combining sam with lim- ited data for change detection in remote sensing. IEEE Transactions on Geoscience and Remote Sensing (2025)

    2025

  8. [8]

    IEEE Transactions on Knowledge and Data Engineering (2025)

    Guan, R., Liu, T., Tu, W., Tang, C., Luo, W., Liu, X.: Sampling enhanced con- trastive multi-view remote sensing data clustering with long-short range informa- tion mining. IEEE Transactions on Knowledge and Data Engineering (2025)

    2025

Show all 80 references
  1. [9]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Hu, Y., Wang, Q., Shao, W., Xie, E., Li, Z., Han, J., Luo, P.: Beyond one-to-one: Rethinking the referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4067–4077 (2023)

    2023

  2. [10]

    In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

    Hu, Z., Feng, G., Sun, J., Zhang, L., Lu, H.: Bi-directional relationship inferring network for referring image segmentation. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 4424–4433 (2020)

    2020

  3. [11]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Huang, J., Xu, Z., Liu, T., Liu, Y., Han, H., Yuan, K., Li, X.: Densely connected parameter-efficient tuning for referring image segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 3653–3661 (2025)

    2025

  4. [12]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Huang, S., Hui, T., Liu, S., Li, G., Wei, Y., Han, J., Liu, L., Li, B.: Referring image segmentation via cross-modal progressive comprehension. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10488–10497 (2020)

    2020

  5. [13]

    In: European conference on computer vision

    Hui, T., Liu, S., Huang, S., Li, G., Yu, S., Zhang, F., Han, J.: Linguistic structure guided context modeling for referring image segmentation. In: European conference on computer vision. pp. 59–75. Springer (2020)

    2020

  6. [14]

    In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Ito, K.: Feature design for bridging sam and clip toward referring image segmenta- tion. In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 8368–8378. IEEE (2025)

    2025

  7. [15]

    In: International conference on machine learning

    Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904–4916. PMLR (2021)

    2021

  8. [16]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015–4026 (2023) CROSS 17

    2023

  9. [17]

    In: International conference on machine learning

    Kornblith, S., Norouzi, M., Lee, H., Hinton, G.: Similarity of neural network rep- resentations revisited. In: International conference on machine learning. pp. 3519–

  10. [18]

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

    Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning seg- mentation via large language model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9579–9589 (2024)

    2024

  11. [19]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–13 (2024)

    Lan, M., Rong, F., Jiao, H., Gao, Z., Zhang, L.: Language query-based transformer with multiscale cross-modal alignment for visual grounding on remote sensing im- ages. IEEE Transactions on Geoscience and Remote Sensing62, 1–13 (2024)

    2024

  12. [20]

    Lei,S.,Xiao,X.,Zhang,T.,Li,H.C.,Shi,Z.,Zhu,Q.:Exploringfine-grainedimage- textalignmentforreferringremotesensingimagesegmentation.IEEETransactions on Geoscience and Remote Sensing63, 1–11 (2025).https://doi.org/10.1109/ TGRS.2024.3522293

    2025

  13. [21]

    In: Proceedings of the AAAI Con- ference on Artificial Intelligence

    Li, B., Dong, H., Zhang, D., Zhao, Z., Sun, H., Gao, J.: Exploring efficient open- vocabulary segmentation in the remote sensing. In: Proceedings of the AAAI Con- ference on Artificial Intelligence. vol. 40, pp. 5982–5991 (2026)

    2026

  14. [22]

    arXiv preprint arXiv:2604.15652 (2026)

    Li, B., Huo, T., Dong, H., Zhang, D., Zhao, Z., Gao, J., Li, X.: Towards realis- tic open-vocabulary remote sensing segmentation: Benchmark and baseline. arXiv preprint arXiv:2604.15652 (2026)

    2026 arXiv

  15. [23]

    arXiv preprint arXiv:2606.00987 (2026)

    Li, B., Zhang, D., Huo, T., Zhao, Z., Gao, J., Li, X.: An open-source bench- mark and baseline for multi-temporal referring segmentation. arXiv preprint arXiv:2606.00987 (2026)

    2026 arXiv

  16. [24]

    In: Proceedings of the 33rd ACM International Conference on Multimedia

    Li, B., Zhang, D., Zhao, Z., Gao, J., Li, X.: Stitchfusion: Weaving any visual modalities to enhance multimodal semantic segmentation. In: Proceedings of the 33rd ACM International Conference on Multimedia. pp. 1308–1317 (2025)

    2025

  17. [25]

    Pattern Recognition168, 111801 (2025)

    Li, B., Zhang, D., Zhao, Z., Gao, J., Li, X.: U3m: Unbiased multiscale modal fusion model for multimodal semantic segmentation. Pattern Recognition168, 111801 (2025)

    2025

  18. [26]

    arXiv preprint arXiv:2504.09644 (2025)

    Li, K., Xin, Z., Pang, L., Pang, C., Deng, Y., Yao, J., Xia, G., Meng, D., Wang, Z., Cao, X.: Segearth-r1: Geospatial pixel reasoning via large language model. arXiv preprint arXiv:2504.09644 (2025)

    2025 arXiv

  19. [27]

    ISPRS Journal of Photogrammetry and Remote Sensing226, 350–363 (2025)

    Li, K., Vosselman, G., Yang, M.Y.: Scale-wise bidirectional alignment network for referring remote sensing image segmentation. ISPRS Journal of Photogrammetry and Remote Sensing226, 350–363 (2025)

    2025

  20. [28]

    International Journal of Applied Earth Observation and Geoinformation124, 103497 (2023)

    Li, X., Wen, C., Hu, Y., Zhou, N.: Rs-clip: Zero shot remote sensing scene classifica- tion via contrastive vision-language supervision. International Journal of Applied Earth Observation and Geoinformation124, 103497 (2023)

    2023

  21. [29]

    arXiv preprint arXiv:2406.00670 (2024)

    Li, Y., Li, Z., Zeng, Q., Hou, Q., Cheng, M.M.: Cascade-clip: Cascaded vision- language embeddings alignment for zero-shot semantic segmentation. arXiv preprint arXiv:2406.00670 (2024)

    2024 arXiv

  22. [30]

    arXiv preprint arXiv:2506.03147 (2025)

    Lin, B., Li, Z., Cheng, X., Niu, Y., Ye, Y., He, X., Yuan, S., Yu, W., Wang, S., Ge, Y., et al.: Uniworld-v1: High-resolution semantic encoders for unified visual understanding and generation. arXiv preprint arXiv:2506.03147 (2025)

    2025 arXiv

  23. [31]

    In: Proceedings of the IEEE international conference on computer vision

    Liu, C., Lin, Z., Shen, X., Yang, J., Lu, X., Yuille, A.: Recurrent multimodal inter- action for referring image segmentation. In: Proceedings of the IEEE international conference on computer vision. pp. 1271–1280 (2017)

    2017

  24. [32]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–16 (2024)

    Liu, F., Chen, D., Guan, Z., Zhou, X., Zhu, J., Ye, Q., Fu, L., Zhou, J.: Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing62, 1–16 (2024)

    2024

  25. [33]

    Advances in neural information processing systems36, 34892–34916 (2023) 18 T

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems36, 34892–34916 (2023) 18 T. Luo et al

    2023

  26. [34]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2025)

    Liu, M., Jiang, X., Zhang, X.: Cadformer: Fine-grained cross-modal alignment and decoding transformer for referring remote sensing image segmentation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2025)

    2025

  27. [35]

    IEEE Transac- tions on Geoscience and Remote Sensing62, 1–15 (2024)

    Liu, R., Luo, T., Huang, S., Wu, Y., Jiang, Z., Zhang, H.: Crossmatch: Cross-view matching for semi-supervised remote sensing image segmentation. IEEE Transac- tions on Geoscience and Remote Sensing62, 1–15 (2024)

    2024

  28. [36]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(9), 4761–4775 (2021)

    Liu, S., Hui, T., Huang, S., Wei, Y., Li, B., Li, G.: Cross-modal progressive com- prehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence44(9), 4761–4775 (2021)

    2021

  29. [37]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liu, S., Ma, Y., Zhang, X., Wang, H., Ji, J., Sun, X., Ji, R.: Rotated multi-scale interaction network for referring remote sensing image segmentation. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26658–26668 (2024)

    2024

  30. [38]

    In: Proceedings of the 31st ACM International Con- ference on Multimedia

    Liu, S.A., Zhang, Y., Qiu, Z., Xie, H., Zhang, Y., Yao, T.: Caris: Context-aware referring image segmentation. In: Proceedings of the 31st ACM International Con- ference on Multimedia. pp. 779–788 (2023)

    2023

  31. [39]

    arXiv preprint arXiv:2503.06520 (2025)

    Liu, Y., Peng, B., Zhong, Z., Yue, Z., Lu, F., Yu, B., Jia, J.: Seg-zero: Reasoning-chain guided segmentation via cognitive reinforcement. arXiv preprint arXiv:2503.06520 (2025)

    2025 arXiv

  32. [40]

    In: 2024 IEEE International Conference on Image Processing (ICIP)

    Luo, T., Du, M., Shi, J., Chen, X., Zhao, B., Huang, S.: Contextuality helps repre- sentation learning for generalized category discovery. In: 2024 IEEE International Conference on Image Processing (ICIP). pp. 687–693. IEEE (2024)

    2024

  33. [41]

    arXiv preprint arXiv:2409.20098 (2024)

    Luo, T., Liu, Y., Liu, Y., Zhang, A., Wang, X., Zhan, Y., Tang, C., Liu, L., Chen, Z.: Dig-face: De-biased learning for generalized facial expression category discovery. arXiv preprint arXiv:2409.20098 (2024)

    2024 arXiv

  34. [42]

    IEEE Transactions on Multimedia (2026)

    Luo, T., Liu, Y., Xu, Y., Liu, R., Wang, X., Zeng, H., Huang, S., Zhang, H.: Stroke-based perception: Discover novel oracle characters. IEEE Transactions on Multimedia (2026)

    2026

  35. [43]

    IEEE Transactions on Geoscience and Remote Sensing (2025)

    Ma, Q., Li, L., Lu, X., Jiao, L., Liu, F., Ma, W., Liu, X., Sun, L.: Lscf: Long-term semantic-guidance convformer for referring remote sensing image segmentation. IEEE Transactions on Geoscience and Remote Sensing (2025)

    2025

  36. [44]

    ISPRS Journal of Photogrammetry and Remote Sensing221, 44–63 (2025)

    Ma, X., Lian, R., Wu, Z., Guan, R., Hong, T., Zhao, M., Ma, M., Nie, J., Du, Z., Song, S., et al.: A novel scene coupling semantic mask network for remote sensing image segmentation. ISPRS Journal of Photogrammetry and Remote Sensing221, 44–63 (2025)

    2025

  37. [45]

    In: Proceedings of the International Conference on 3D Vision (3DV)

    Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: Proceedings of the International Conference on 3D Vision (3DV). pp. 565–571 (2016)

    2016

  38. [46]

    International Journal of Applied Earth Observation and Geoinformation124, 103540 (2023)

    Osco, L.P., Wu, Q., De Lemos, E.L., Gonçalves, W.N., Ramos, A.P.M., Li, J., Junior, J.M.: The segment anything model (sam) for remote sensing applications: From zero to one shot. International Journal of Applied Earth Observation and Geoinformation124, 103540 (2023)

    2023

  39. [47]

    IEEE Geoscience and Remote Sensing Magazine (2025)

    Ou, R., Hu, Y., Zhang, F., Chen, J., Liu, Y.: Geopix: A multimodal large language model for pixel-level image understanding in remote sensing. IEEE Geoscience and Remote Sensing Magazine (2025)

    2025

  40. [48]

    In: Proceedings of the 32nd ACM International Conference on Multi- media

    Pan, Y., Sun, R., Wang, Y., Zhang, T., Zhang, Y.: Rethinking the implicit opti- mization paradigm with dual alignments for referring remote sensing image seg- mentation. In: Proceedings of the 32nd ACM International Conference on Multi- media. pp. 2031–2040 (2024) CROSS 19

    2024

  41. [49]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  42. [50]

    arXiv preprint arXiv:2408.00714 (2024)

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., et al.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

    2024 arXiv

  43. [51]

    arXiv e-prints pp

    Rong, F., Lan, M., Zhang, Q., Zhang, L.: Customized sam 2 for referring remote sensing image segmentation. arXiv e-prints pp. arXiv–2503 (2025)

    2025

  44. [52]

    arXiv preprint arXiv:2501.13925 (2025)

    Shabbir, A., Zumri, M., Bennamoun, M., Khan, F.S., Khan, S.: Geopixel: Pixel grounding large multimodal model in remote sensing. arXiv preprint arXiv:2501.13925 (2025)

    2025 arXiv

  45. [53]

    In: Proceedings of the 30th ACM International conference on Multimedia

    Sun, Y., Feng, S., Li, X., Ye, Y., Kang, J., Huang, X.: Visual grounding in re- mote sensing images. In: Proceedings of the 30th ACM International conference on Multimedia. pp. 404–412 (2022)

    2022

  46. [54]

    IEEE Transactions on Geoscience and Remote Sensing (2026)

    Sun, Y., Dong, Z., Jiang, H., Gu, Y., Liu, T.: Crobim-u: Uncertainty-driven re- ferring remote sensing image segmentation. IEEE Transactions on Geoscience and Remote Sensing (2026)

    2026

  47. [55]

    arXiv preprint arXiv:2502.14786 (2025)

    Tschannen, M., Gritsenko, A., Wang, X., Naeem, M.F., Alabdulmohsin, I., Parthasarathy, N., Evans, T., Beyer, L., Xia, Y., Mustafa, B., et al.: Siglip 2: Multilingual vision-language encoders with improved semantic understanding, lo- calization, and dense features. arXiv prepri...

    2025 arXiv

  48. [56]

    arXiv preprint arXiv:2509.02273 (2025)

    Wang, C., Ji, Y., Meng, Y., Zhang, Y., Zhu, Y.: Rs-ood: A vision-language augmented framework for out-of-distribution detection in remote sensing. arXiv preprint arXiv:2509.02273 (2025)

    2025

  49. [57]

    arXiv preprint arXiv:2509.03002 (2025)

    Wang, C., Ji, Y., Meng, Y., Zhang, Y., Zhu, Y.: Sopseg: Prompt-based small object instance segmentation in remote sensing imagery. arXiv preprint arXiv:2509.03002 (2025)

    2025 arXiv

  50. [58]

    IEEE Transactions on Geoscience and Remote Sensing (2025)

    Wang, C., Xi, Z., Liu, D., Feng, Y., Deng, Y., Li, K., Chen, J., Chen, J., Meng, Y.: Pcp: A prompt-based cartographic-level polygonal vector extraction framework for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing (2025)

    2025

  51. [59]

    Advances in Neural Information Processing Systems36, 8815–8827 (2023)

    Wang, D., Zhang, J., Du, B., Xu, M., Liu, L., Tao, D., Zhang, L.: Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems36, 8815–8827 (2023)

    2023

  52. [60]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, F., Liu, H.: Understanding the behaviour of contrastive loss. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2495–2504 (2021)

    2021

  53. [61]

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

    Wang, J., Ke, L.: Llm-seg: Bridging image segmentation and large language model reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1765–1774 (2024)

    2024

  54. [62]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: Cris: Clip-driven referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11686–11695 (2022)

    2022

  55. [63]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wang, Z., Prabha, R., Huang, T., Wu, J., Rajagopal, R.: Skyscript: A large and semantically diverse vision-language dataset for remote sensing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 5805–5813 (2024)

    2024

  56. [64]

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

    Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni...

    2024

  57. [65]

    arXiv preprint arXiv:2311.05348 (2023)

    Xu, J., Xu, L., Yang, Y., Li, X., Wang, F., Xie, Y., Huang, Y.J., Li, Y.: u-llava: Uni- fying multi-modal tasks via large language model. arXiv preprint arXiv:2311.05348 (2023)

    2023 arXiv

  58. [66]

    In: Proceedings of the 2026 Inter- national Conference on Multimedia Retrieval

    Xu, Y., Guo, C., Shuai, Y., Ni, J.: Relational retrieval: Leveraging known-novel interactions for generalized category discovery. In: Proceedings of the 2026 Inter- national Conference on Multimedia Retrieval. pp. 410–414 (2026)

    2026

  59. [67]

    IEEE Transactions on Geoscience and Remote Sensing61, 1–16 (2023)

    Yan, Z., Li, J., Li, X., Zhou, R., Zhang, W., Feng, Y., Diao, W., Fu, K., Sun, X.: Ringmo-sam: A foundation model for segment anything in multimodal remote- sensing images. IEEE Transactions on Geoscience and Remote Sensing61, 1–16 (2023)

    2023

  60. [68]

    In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition

    Yang, K., Wang, Y., Luo, T.: Assignment-driven hash learning in a hyper-semantic space for on-the-fly category discovery. In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition. pp. 11303–11312 (2026)

    2026

  61. [69]

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

    Yang, S., Xia, M., Li, G., Zhou, H.Y., Yu, Y.: Bottom-up shift and reasoning for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11266–11275 (2021)

    2021

  62. [70]

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

    Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavt: Language- aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18155– 18165 (2022)

    2022

  63. [71]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–12 (2024)

    Yuan, Z., Mou, L., Hua, Y., Zhu, X.X.: Rrsis: Referring remote sensing image segmentation. IEEE Transactions on Geoscience and Remote Sensing62, 1–12 (2024)

    2024

  64. [72]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Yue, W., Zhang, J., Hu, K., Xia, Y., Luo, J., Wang, Z.: Surgicalsam: Efficient class promptable surgical instrument segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6890–6898 (2024)

    2024

  65. [73]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language im- age pre-training. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 11975–11986 (2023)

    2023

  66. [74]

    IEEE Transactions on Geoscience and Remote Sensing 61, 1–13 (2023)

    Zhan, Y., Xiong, Z., Yuan, Y.: Rsvg: Exploring data and models for visual ground- ing on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 61, 1–13 (2023)

    2023

  67. [75]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–20 (2024)

    Zhang, W., Cai, M., Zhang, T., Zhuang, Y., Mao, X.: Earthgpt: A universal multi- modallargelanguagemodelformultisensorimagecomprehensioninremotesensing domain. IEEE Transactions on Geoscience and Remote Sensing62, 1–20 (2024)

    2024

  68. [76]

    IEEE Transactions on Image Processing (2025)

    Zhang, X., Zhang, H., Wang, G., Zhang, Q., Zhang, L., Du, B.: Uniuir: Considering underwater image restoration as an all-in-one learner. IEEE Transactions on Image Processing (2025)

    2025

  69. [77]

    arXiv preprint arXiv:2406.20076 (2024)

    Zhang, Y., Cheng, T., Hu, R., Liu, L., Liu, H., Ran, L., Chen, X., Liu, W., Wang, X.: Evf-sam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:2406.20076 (2024)

    2024 arXiv

  70. [78]

    ISPRS Journal of Photogrammetry and Remote Sensing237, 323–338 (2026)

    Zhang, Z., Li, J., Liang, Y., Yan, J., Xiao, Y., Su, X., Yuan, Q.: Ecrformer: An effi- cient cloud removal transformer with semantic-decoupled learning for multimodal satellite imagery. ISPRS Journal of Photogrammetry and Remote Sensing237, 323–338 (2026)

    2026

  71. [79]

    IEEE Transactions on Geoscience and Remote Sensing (2025)

    Zhang, Z., Yan, J., Liang, Y., Feng, J., He, H., Cao, L.: Multi-scale restoration of missing data in optical time-series images with masked spatial-temporal attention network. IEEE Transactions on Geoscience and Remote Sensing (2025)

    2025

  72. [80]

    A baseball field at thebottom

    Zhong, Z., Tang, Z., He, T., Fang, H., Yuan, C.: Convolution meets lora: Parameter efficient finetuning for segment anything model. arXiv preprint arXiv:2401.17868 (2024) CROSS 21 Appendix 1 Spatial Counterfactual Generation: Rules, Cases, and LLM Ablation This section provide...

    2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.