Pith. sign in

REVIEW 6 major objections 5 minor 48 references

Representation Discrepancy Bridging Method for Remote Sensing Image-Text Retrieval

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read An asymmetric adapter with a dual-task loss outperforms full fine-tuning on remote-sensing image-text retrieval.

desk verdict A plausible adapter-based PEFT method for remote sensing retrieval, but the headline comparison against PEFT baselines is confounded by a different backbone, and the same-backbone gain lacks variance. read the letter →

arxiv 2505.16756 v1 pith:V4PO4DFF submitted 2025-05-22 cs.CV cs.IRcs.MM

Hailong Ning , Siying Wang , Tao Lei , Xiaopeng Cao , Huanmin Dou , Bin Zhao , Asoke K. Nandi , Petia Radeva This is my paper
classification cs.CVcs.IRcs.MM
keywords RemoteSensingImage-TextRetrievalVision-and-LanguagePre-trainingParameter-EfficientFine-TuningCross-ModalAsymmetricAdapterDual-TaskConsistencyLossDifferentialAttentionHierarchicalMulti-TaskLearning
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

Remote sensing image-text retrieval pairs satellite or aerial images with natural-language captions, and the paper argues that existing parameter-efficient fine-tuning methods for vision-language models are held back by a specific failure: the text side is so class-discriminative that it dominates the shared optimization and prevents the image encoder from learning fine-grained visual details. To close that gap, the Representation Discrepancy Bridging (RDB) method uses a Cross-Modal Asymmetric Adapter (CMAA), with a Visual Enhancement Adapter based on differential attention for images and a Text Semantic Adapter based on hierarchical attention for text, plus a Dual-Task Consistency Loss (DTCL) that combines cross-modal ranking, classification, and exponential-moving-average consistency terms with automated weighting. On the RSICD and RSITMD benchmarks, RDB raises mean recall to 38.09 and 52.04, beating the fully fine-tuned GeoRSCLIP model (36.94 and 50.14) while updating only adapter parameters. If the results hold, parameter-efficient asymmetric adapters plus auxiliary losses are a viable substitute for full fine-tuning in this domain.

What carries the argument

Two coupled mechanisms carry the argument. The first is the Cross-Modal Asymmetric Adapter (CMAA), in which a Visual Enhancement Adapter (VEA) computes differential attention, $(\mathrm{softmax}(Q_1K_1^\top/\sqrt{d}) - \lambda\,\mathrm{softmax}(Q_2K_2^\top/\sqrt{d}))V$ with a learnable $\lambda$, in order to sharpen fine-grained image features, while a Text Semantic Adapter (TSA) runs hierarchical attention with word- and sentence-level context vectors over a bi-directional GRU to emphasize informative words; a shared gated layer allows the two branches to exchange information. The second is the Dual-Task Consistency Loss (DTCL), which sums a margin-based cross-modal ranking loss, a cross-entropy classification loss, and an exponential-moving-average consistency loss that treats the text encoder as a teacher and the image encoder as a student, all combined with uncertainty-based learnable weights $\sigma_i$. The asymmetry itself is the load-bearing idea: the text branch needs only to select existing semantics, while the image branch needs an enhanced pathway to recover detail.

What would settle it

Train the same frozen backbone with a symmetric adapter and with the asymmetric CMAA on a dataset where the captions have been made as hard to separate by class as the images (for example, by injecting label noise or shuffling words). If the asymmetric method no longer beats the symmetric one, the imbalance premise is what drives the gain. A direct alternative is to measure class separability of image versus text embeddings under a symmetric adapter: the paper's mechanism predicts a clear gap, and the improvement from asymmetry should scale with that gap.

Watch

Extended reading notes

Core claim

The central claim is that the bottleneck in parameter-efficient fine-tuning of vision-language models for remote sensing is not parameter count but an imbalanced cross-modal optimization: because text features are naturally more separable by class than image features, a symmetric adapter lets the text modality dominate the contrastive objective and stalls image representation learning. The RDB method responds with two coordinated changes. First, it makes the adapter architecture asymmetric: the Visual Enhancement Adapter applies differential attention—the difference of two softmax attention maps with a learnable mixing scalar $\lambda$—to draw out fine-grained image cues, while the Text Semantic Adapter uses hierarchical word- and sentence-level attention to concentrate on informative keywords; the two branches meet in shared gated layers. Second, it replaces the single cross-modal loss with a dual-task objective, DTCL, that adds a classification term and an EMA-based teacher-student consistency term to the standard margin-based alignment loss, with uncertainty-based weights that adapt during training. On RSICD and RSITMD, this configuration achieves mean recall of 38.09 and 52.04, surpassing the fully fine-tuned GeoRSCLIP at 36.94 and 50.14 and all compared PEFT methods, which the paper attributes to the asymmetry and the multi-task consistency rather than to additional capacity.

Load-bearing premise

The load-bearing premise is that text features are already much more class-separable than image features in the shared embedding space, so a symmetric adapter lets the text side dominate and hold back the image side; this premise comes from a qualitative 2-D visualization, not a measured statistic, and if the imbalance is absent the asymmetric design has no stated reason to outperform a symmetric one.

Editorial extensions

If this is right

  • Parameter-efficient fine-tuning can reach beyond full fine-tuning in remote sensing image-text retrieval, so practitioners can keep the backbone frozen and still improve over re-training all weights.
  • The dual-task loss recipe—cross-modal ranking plus classification plus EMA consistency with adaptive weighting—should transfer to other cross-modal retrieval settings where one modality is easier to discriminate than the other.
  • Differential attention in the image adapter gives a concrete way to extract fine-grained cues such as object counts and small structures that full fine-tuning misses, as shown in the qualitative retrieval examples.
  • Because only adapter parameters are updated, the method is cheaper to train and less prone to catastrophic forgetting than full fine-tuning, making it more practical for repeated domain updates.

Reading between the lines

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

  • A quantitative class-separability measurement, such as the ratio of between-class to within-class cosine distances for image and text embeddings, would convert the paper's motivating visualization into a testable predictor of when asymmetric adapters help.
  • The EMA consistency term effectively distills text-side semantics into the image encoder; a natural extension outside retrieval would be image-only classification that borrows the same teacher-student term.
  • The reported gains are on two remote sensing datasets; the same asymmetry logic may apply in other domains with concise captions such as medical imaging or industrial inspection, but the paper does not test that transfer.
  • Because the advantage over full fine-tuning is 1.15–2%, isolating the Visual Enhancement Adapter from the Text Semantic Adapter and testing across backbone sizes would show how much of the gain each branch contributes.
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

6 major / 5 minor

Summary. The manuscript proposes the Representation Discrepancy Bridging (RDB) method for remote sensing image-text retrieval. RDB combines a Cross-Modal Asymmetric Adapter (CMAA), composed of a Visual Enhancement Adapter (VEA) with differential attention, a Text Semantic Adapter (TSA) with hierarchical attention, and shared/gated interaction layers, with a Dual-Task Consistency Loss (DTCL) that jointly optimizes a cross-modal margin loss, a classification loss, and an EMA-based consistency loss with learnable weights. The method is evaluated on RSICD and RSITMD using a GeoRSCLIP backbone, with 5-fold cross-validation and reporting R@K and mR. The main claims are that RDB improves mR by 1.15%-2% over full fine-tuned GeoRSCLIP and by 6%-11% over state-of-the-art PEFT methods.

Significance. The idea of using asymmetric image/text adapters and combining cross-modal, classification, and consistency objectives is timely and potentially useful for remote sensing retrieval. If validated, the paper would offer a parameter-efficient alternative that surpasses full fine-tuning, which is an important practical result. The paper has several strengths: it uses standard benchmark datasets, reports results for all R@K and mR metrics, includes an ablation of the three DTCL loss terms (Table 3), and provides qualitative retrieval examples. However, the current evidence does not establish the headline claims. The PEFT comparison is confounded by different backbones, no variance or significance testing is reported, and the mechanism underlying the asymmetric design is only supported by a qualitative t-SNE plot. The significance is therefore conditional on the additional experiments requested below.

major comments (6)
  1. [Table 1] Table 1: The PEFT comparison is confounded by backbone choice. All PEFT baselines (Adapter, CLIP-Adapter, AdaptFormer, Cross-Modal Adapter, UniAdapter, PE-RSITR) are evaluated on CLIP(ViT-B-32), while RDB uses GeoRSCLIP(ViT-B-32-RET-2). The Full-FT rows show that the backbone shift alone accounts for a large share of the reported margin: Full-FT GeoRSCLIP beats Full-FT CLIP by 4.61 mR on RSICD (36.94 vs 32.33) and by 4.01 mR on RSITMD (50.14 vs 46.13), whereas RDB's margins over PE-RSITR are 6.97 and 7.57 mR, respectively. To support the claim that RDB "surpasses existing PEFT methods," the paper needs same-backbone PEFT baselines, e.g., PE-RSITR, CLIP-Adapter, and AdaptFormer trained on GeoRSCLIP.
  2. [§4.2, Tables 1-3] All quantitative results are reported as 5-fold averages without standard deviations, per-fold numbers, or significance tests. The central same-backbone gains are small: 1.15 mR on RSICD (38.09 vs 36.94) and 1.90 mR on RSITMD (52.04 vs 50.14). Without an estimate of fold-to-fold variability, the reader cannot determine whether these differences are systematic or within noise. Please report standard deviations or per-fold results for at least the RDB, Full-FT GeoRSCLIP, and key baseline rows, and perform a paired test where possible.
  3. [§1, Figure 1; Table 2] The premise for the asymmetric design—that text features' higher inter-class discriminability dominates the cross-modal optimization process and inhibits image representation learning—rests on a qualitative t-SNE visualization. No quantitative measure (e.g., silhouette score, kNN classification accuracy, or feature-norm statistics) is given, and the ablation in Table 2 only adds the full CMAA, never isolating VEA versus TSA or comparing with a symmetric adapter of matched capacity. Without such evidence, the reported mR gains do not distinctly support the claimed mechanism. Please add a quantitative discriminability comparison and a symmetric-adapter ablation.
  4. [§3.2, Figure 3] The CMAA's "shared layer" and "gating mechanism" are central architectural components but are not defined. There are no equations for the shared-layer transformation, the gated sum, or the gating weights; it is also unclear whether parameters are shared across the two adapters or within each adapter, and at which Transformer blocks the adapters are inserted. These details are necessary for reproducibility and for isolating the effect of the shared/gating components.
  5. [§3.3, Eq. (13)] The consistency constraint is described as an EMA teacher-student framework (text teacher, image student), but the EMA update rule, the stop-gradient behavior, and the exact definitions of y_Ii and y_Ti are not specified. Eq. (13) is written as a plain MSE between image and text outputs, which leaves ambiguity about whether the same projection space is used and whether the text branch is detached. Also, Table 3 shows that the consistency term in isolation (Method 3, mR 51.43) improves less than the classification term (Method 2, mR 51.68), so the claim that consistency "further alleviates" the imbalance should be treated cautiously.
  6. [Section 4] The paper does not report the number of trainable parameters for RDB, full fine-tuning, or the PEFT baselines. Since the method is presented as parameter-efficient, the absence of parameter counts makes it impossible to assess this core claim. Please include trainable parameter counts and, ideally, training cost metrics such as GPU hours or memory usage.
minor comments (5)
  1. [§3.2] The phrase "HIA ensures efficient textual feature learning" appears to be a typo; it should be "HA" (Hierarchical Attention).
  2. [§4.2] "a weight decay of 0.7 was applied after every 20 training epochs" should presumably say "learning rate decay" rather than "weight decay."
  3. [Eq. (14)] The regularization term log(1+sigma_i^2) in Eq. (14) is unconventional; the standard formulation in the cited multi-task learning literature uses log(sigma_i^2). Please justify this choice or correct the formula.
  4. [Table 1] The table would be easier to read if a separate column explicitly labeled the pretraining dataset/backbone variant for every row, since the current "Backbone" column mixes CLIP and GeoRSCLIP without highlighting the confounding influence.
  5. [Abstract and Introduction] There are minor typos, including "modalites" and "lightw-eight group convolution," which should be corrected in a final pass.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: method components and losses are not defined in terms of the target metric; self-citation [5] is background only.

full rationale

The derivation chain is self-contained: RDB is a proposed architecture (CMAA with VEA/TSA) plus a composite loss (DTCL), evaluated against external baselines on RSICD and RSITMD. No component of the method is defined in terms of the reported metric mR, no fitted parameter is relabeled as a prediction, and no load-bearing premise is justified solely by the authors' prior work. The only self-citation, [5], appears in the introduction as background support for the general claim that cross-modal semantic alignment is a core challenge, and it plays no role in deriving the adapter or the loss. The central comparison to full fine-tuned GeoRSCLIP uses the same backbone and is an external benchmark, though the PEFT comparison in Table 1 mixes backbones (CLIP ViT-B-32 vs GeoRSCLIP ViT-B-32-RET-2) and is reported without variance; that is a correctness and experimental-design concern about isolating the method's contribution, not a circularity in which an output reduces to an input. Potential weaknesses such as the qualitative t-SNE premise and the absence of same-backbone PEFT baselines are evidentiary or design limitations, not self-referential reductions. Accordingly, no circular step is exhibited.

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

No new physical or conceptual entities are introduced; all components are borrowed neural modules (differential attention, hierarchical attention, EMA) and loss terms. The central claim rests on standard deep learning training plus the empirical premise that text discriminability dominates and that the borrowed attention mechanisms fix it.

free parameters (4)
  • Cross-modal margin µ = 0.2
    Boundary in Eq. (10), set empirically in Section 4.2; no sensitivity analysis shown.
  • λinit = constant in (0,1), exact value not reported
    Initialization of the learnable differential attention scalar in Eq. (3); chosen by hand.
  • Learnable loss weights σ_i = learned during training
    Weights in Eq. (14) combine the three loss terms; fitted on training data rather than fixed.
  • Dropout rate = 0.2
    Chosen empirically in Section 4.2; not reported as a tuned hyperparameter.
assumptions (4)
  • domain assumption Text modality's high inter-class discriminability dominates cross-modal optimization and inhibits image learning.
    Motivates the asymmetric design; supported only by qualitative t-SNE in Figure 1, not by a quantitative metric.
  • domain assumption Differential attention extracts better fine-grained image features in remote sensing.
    Borrowed from [18], applied to RS images; no isolated VEA ablation in Tables 2 or 3.
  • domain assumption Hierarchical attention identifies key text semantics.
    Borrowed from [19]; no isolated TSA ablation is provided.
  • domain assumption EMA teacher-student consistency aligns modalities without hurting discrimination.
    Standard technique, but no analysis or hyperparameter such as decay is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Representation Discrepancy Bridging Method for Remote Sensing Image-Text Retrieval." pith.science (2026). https://pith.science/paper/V4PO4DFF

@misc{pith2026250516756,
  author       = {Pith},
  title        = {Pith review of: Representation Discrepancy Bridging Method for Remote Sensing Image-Text Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V4PO4DFF}},
  note         = {Machine review of arXiv:2505.16756}
}
read the original abstract

Remote Sensing Image-Text Retrieval (RSITR) plays a critical role in geographic information interpretation, disaster monitoring, and urban planning by establishing semantic associations between image and textual descriptions. Existing Parameter-Efficient Fine-Tuning (PEFT) methods for Vision-and-Language Pre-training (VLP) models typically adopt symmetric adapter structures for exploring cross-modal correlations. However, the strong discriminative nature of text modality may dominate the optimization process and inhibits image representation learning. The nonnegligible imbalanced cross-modal optimization remains a bottleneck to enhancing the model performance. To address this issue, this study proposes a Representation Discrepancy Bridging (RDB) method for the RSITR task. On the one hand, a Cross-Modal Asymmetric Adapter (CMAA) is designed to enable modality-specific optimization and improve feature alignment. The CMAA comprises a Visual Enhancement Adapter (VEA) and a Text Semantic Adapter (TSA). VEA mines fine-grained image features by Differential Attention (DA) mechanism, while TSA identifies key textual semantics through Hierarchical Attention (HA) mechanism. On the other hand, this study extends the traditional single-task retrieval framework to a dual-task optimization framework and develops a Dual-Task Consistency Loss (DTCL). The DTCL improves cross-modal alignment robustness through an adaptive weighted combination of cross-modal, classification, and exponential moving average consistency constraints. Experiments on RSICD and RSITMD datasets show that the proposed RDB method achieves a 6%-11% improvement in mR metrics compared to state-of-the-art PEFT methods and a 1.15%-2% improvement over the full fine-tuned GeoRSCLIP model.

Figures

Figures reproduced from arXiv: 2505.16756 by the authors.

Figure 1
Figure 1. The t-SNE visualization of features from text modality (left panel) and image modality (right panel) when utilizing a symmetric adapter structure for the RSITR task. methods, adapter [17] is the most commonly adopted fine-tuning scheme. However, in the RSITR task, the structural design of adapter typically employs symmetric architec￾ture for both image and text modalities. This design ignores the inherent difference… view at source ↗
Figure 2
Figure 2. The overall framework of the RDB method. which may lead to imbalanced cross-modal optimization problem. To address this lim￾itation, a Cross-Modal Asymmetric Adapter (CMAA) architecture is designed to fa￾cilitate modality-specific optimization and improve feature alignment. Meanwhile, the DTCL loss function is proposed to extend the traditional single-task retrieval model into a multi-task optimization model for eff… view at source ↗
Figure 3
Figure 3. (a) The DA mechanism introduced in the VEA. (b) The specific structure of the CMAA. (c) The HA mechanism introduced in the TSA. cient textual feature learning. In the CMAA module, the feature extraction pipeline operates through three sequential phases. First, the image feature EI and text feature ET through Transformer blocks undergo a dimensionality reduction via linear projec￾tion, followed by GELU activation fun… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of Top-5 results between the proposed RDB method and the Full-FT GeoRSCLIP method on the RSITMD dataset for the image-text retrieval task. Red markings indicate retrieval errors; the last column shows the correct RS image corresponding to the incor￾rectly re…
Figure 5
Figure 5. Figure 5: Comparison of Top-5 results between the proposed RDB method and the Full-FT GeoRSCLIP method in the text-image retrieval task on the RSITMD dataset. The portion marked by the orange box indicates the RS image that matches the retrieved text. These results fully demonst…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 40 canonical work pages

  1. [1]

    Jovhari, A

    N. Jovhari, A. Sedaghat, R. Shah-Hosseini, N. Mohammadi, M. Hasanlou, Noise- robust multimodal remote sensing image matching via geometric analysis of em- bedded pre-trained manifolds, Neurocomputing 638 (2025) 130150

    work page 2025

  2. [2]

    J. Wang, S. Zhang, I. Lizaga, Y . Zhang, X. Ge, Z. Zhang, W. Zhang, Q. Huang, Z. Hu, Uas-based remote sensing for agricultural monitoring: Current status and perspectives, Comput. Electron. Agric. 227 (2024) 109501

    work page 2024

  3. [3]

    B. Dong, R. Zhang, S. Li, Y . Ye, C. Huang, A meta-analysis for the nighttime light remote sensing data applied in urban research: Key topics, hotspot study areas and new trends, Sci. Remote Sensing 11 (2025) 100186

    work page 2025

  4. [4]

    Zhang, W

    Y . Zhang, W. Li, M. Zhang, Y . Qu, R. Tao, H. Qi, Topological structure and semantic information transfer network for cross-scene hyperspectral image clas- sification, IEEE Trans. Neural Netw. Learn. Syst. 34 (2023) 2817–2830

    work page 2023

  5. [5]

    H. Ning, B. Zhao, Y . Yuan, Semantics-consistent representation learning for re- mote sensing image–voice retrieval, IEEE Trans. Geosci. Remote Sens. 60 (2022) 1–14. 23

    work page 2022

  6. [6]

    Abdullah, Y

    T. Abdullah, Y . Bazi, M. M. Al Rahhal, M. L. Mekhalfi, L. Rangarajan, M. Zuair, Textrs: Deep bidirectional triplet network for matching text to remote sensing images, Remote Sens. 12 (2020) 405

    work page 2020

  7. [7]

    Hoxha, F

    G. Hoxha, F. Melgani, B. Demir, Toward remote sensing image retrieval under a deep image captioning perspective, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 13 (2020) 4462–4475

    work page 2020

  8. [8]

    M. M. A. Rahhal, Y . Bazi, T. Abdullah, M. L. Mekhalfi, M. Zuair, Deep unsu- pervised embedding for remote sensing image retrieval using textual cues, Appl. Sci. 10 (2020) 8931

    work page 2020

Show all 48 references
  1. [9]

    Z. Li, X. Wu, H. Du, F. Liu, H. Nghiem, G. Shi, Benchmark evaluations, ap- plications, and challenges of large vision language models: A survey (2025). doi:https://doi.org/10.32388/GXR68Q

    2025 doi

  2. [10]

    S. Lu, J. Guo, J. Zimmer-Dauphinee, J. M. Nieusma, X. Wang, P. VanValken- burgh, S. A. Wernke, Y . Huo, Ai foundation models in remote sensing: A survey, CoRR abs/2408.03464 (2024)

    2024 arXiv

  3. [11]

    M. U. Khattak, H. Rasheed, M. Maaz, S. Khan, F. S. Khan, Maple: Multi-modal prompt learning, in: in Proc. 2023 IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2023, pp. 19113–19122

    2023

  4. [12]

    Y . Yuan, Y . Zhan, Z. Xiong, Parameter-efficient transfer learning for remote sens- ing image–text retrieval, IEEE Trans. Geosci. Remote Sens. 61 (2023) 1–14

    2023

  5. [13]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, I. Sutskever, Learning transferable visual models from natural language supervision, in: in Proc. 38th International Conference on Machine Learning, 2021, ...

    2021

  6. [14]

    J. Li, D. Li, C. Xiong, S. Hoi, Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, in: in Proc. 39th Inter- national Conference on Machine Learning, 2022, pp. 12888–12900. 24

    2022

  7. [15]

    F. Liu, D. Chen, Z. Guan, X. Zhou, J. Zhu, Q. Ye, L. Fu, J. Zhou, Remoteclip: A vision language foundation model for remote sensing, IEEE Trans. Geosci. Remote Sens. 62 (2024) 1–16

    2024

  8. [16]

    Zhang, T

    Z. Zhang, T. Zhao, Y . Guo, J. Yin, Rs5m and georsclip: A large-scale vision- language dataset and a large vision-language model for remote sensing, IEEE Trans. Geosci. Remote Sens. 62 (2024) 1–23

    2024

  9. [17]

    Houlsby, A

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Ges- mundo, M. Attariyan, S. Gelly, Parameter-efficient transfer learning for nlp, in: in Proc. 36th International Conference on Machine Learning, 2019, pp. 2790–2799

    2019

  10. [18]

    T. Ye, L. Dong, Y . Xia, Y . Sun, Y . Zhu, G. Huang, F. Wei, Differential transformer, in: in Proc. 13th International Conference on Learning Representations, 2025

    2025

  11. [19]

    Z. Yang, D. Yang, C. Dyer, X. He, A. Smola, E. Hovy, Hierarchical attention net- works for document classification, in: in Proc. 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2016, pp. 1480–1489

    2016

  12. [20]

    Cheng, Y

    Q. Cheng, Y . Zhou, P. Fu, Y . Xu, L. Zhang, A deep semantic alignment network for the cross-modal image-text retrieval in remote sensing, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 14 (2021) 4284–4297

    2021

  13. [21]

    X. Lu, B. Wang, X. Zheng, X. Li, Exploring models and data for remote sensing image caption generation, IEEE Trans. Geosci. Remote Sens. 56 (2018) 2183– 2195

    2018

  14. [22]

    X. Li, X. Zhang, W. Huang, Q. Wang, Truncation cross entropy loss for remote sensing image captioning, IEEE Trans. Geosci. Remote Sens. 59 (2021) 5246– 5257

    2021

  15. [23]

    T. Wang, X. Xu, Y . Yang, A. Hanjalic, H. T. Shen, J. Song, Matching images and text with multi-modal tensor fusion and re-ranking, in: In Proc. 27th ACM International Conference on Multimedia, 2019, p. 12–20. 25

    2019

  16. [24]

    X. Li, Q. Guo, X. Lu, Spatiotemporal statistics for video quality assessment, IEEE Trans. Image Process 25 (7) (2016) 3329–3342

    2016

  17. [25]

    Z. Yuan, W. Zhang, X. Rong, X. Li, J. Chen, H. Wang, K. Fu, X. Sun, A lightweight multi-scale crossmodal text-image retrieval method in remote sens- ing, IEEE Trans. Geosci. Remote Sens. 60 (2022) 1–19

    2022

  18. [26]

    Z. Yuan, W. Zhang, K. Fu, X. Li, C. Deng, H. Wang, X. Sun, Exploring a fine- grained multiscale method for cross-modal remote sensing image retrieval, IEEE Trans. Geosci. Remote Sens. 60 (2022) 1–19

    2022

  19. [27]

    M. M. A. Rahhal, Y . Bazi, N. A. Alsharif, L. Bashmal, N. Alajlan, F. Melgani, Multilanguage transformer for improved text to remote sensing image retrieval, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 15 (2022) 9115–9126

    2022

  20. [28]

    Z. Wang, J. Yu, A. W. Yu, Z. Dai, Y . Tsvetkov, Y . Cao, Simvlm: Simple visual language model pretraining with weak supervision, https://arxiv.org/ abs/2108.10904 (2022)

    2022 arXiv

  21. [29]

    J. Lin, R. Men, A. Yang, C. Zhou, M. Ding, Y . Zhang, P. Wang, A. Wang, L. Jiang, X. Jia, J. Zhang, J. Zhang, X. Zou, Z. Li, X. Deng, J. Liu, J. Xue, H. Zhou, J. Ma, J. Yu, Y . Li, W. Lin, J. Zhou, J. Tang, H. Yang, M6: A chinese multimodal pretrainer, https://arxiv.org/abs/21...

    2021 arXiv

  22. [30]

    W. Kim, B. Son, I. Kim, Vilt: Vision-and-language transformer without con- volution or region supervision, https://arxiv.org/abs/2102.03334 (2021)

    2021 arXiv

  23. [31]

    J. Lu, D. Batra, D. Parikh, S. Lee, Vilbert: Pretraining task-agnostic visiolin- guistic representations for vision-and-language tasks, https://arxiv.org/ abs/1908.02265 (2019)

    2019 arXiv

  24. [32]

    F. Yu, J. Tang, W. Yin, Y . Sun, H. Tian, H. Wu, H. Wang, Ernie-vil: Knowl- edge enhanced vision-language representations through scene graph, https: //arxiv.org/abs/2006.16934 (2021). 26

    2021 arXiv

  25. [33]

    L. Wang, Z. Tong, B. Ji, G. Wu, Tdn: Temporal difference networks for efficient action recognition, in: in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2021, pp. 1895–1904

    2021

  26. [34]

    Li, Positive-incentive noise, IEEE Trans

    X. Li, Positive-incentive noise, IEEE Trans. Neural Netw. Learn. Syst. (2022)

    2022

  27. [35]

    Lester, R

    B. Lester, R. Al-Rfou, N. Constant, The power of scale for parameter-efficient prompt tuning, in: in Proc. 2021 Conference on Empirical Methods in Natural Language Processing, 2021, pp. 3045–3059

    2021

  28. [36]

    X. L. Li, P. Liang, Prefix-tuning: Optimizing continuous prompts for generation, in: in Proc. 59th Annual Meeting of the Association for Computational Linguis- tics and the 11th International Joint Conference on Natural Language Processing, 2021, pp. 4582–4597

    2021

  29. [37]

    M. Jia, L. Tang, B. C. Chen, C. Cardie, S. Belongie, B. Hariharan, S. N. Lim, Visual prompt tuning, in: in Proc. European Conference on Computer Vision, 2022, p. 709–727

    2022

  30. [38]

    Y . Wu, Y . Yu, Z. Yuan, S. Huang, B. Cai, Apt: Adaptive prefix-tuning on pre- trained models for code intelligence, in: in Proc. 2024 International Joint Confer- ence on Neural Networks, 2024, pp. 1–10

    2024

  31. [39]

    P. Gao, S. Geng, R. Zhang, T. Ma, R. Fang, Y . Zhang, H. Li, Y . Qiao, Clip-adapter: Better vision-language models with feature adapters, Int. J. Comput. Vision 132 (2023) 581–595

    2023

  32. [40]

    Kumar, O

    A. Kumar, O. Irsoy, P. Ondruska, M. Iyyer, J. Bradbury, I. Gulrajani, V . Zhong, R. Paulus, R. Socher, Ask me anything: Dynamic memory networks for natural language processing, in: In Proc. 33rd International Conference on International Conference on Machine Learning, 2016, p....

    2016

  33. [41]

    Sukhbaatar, A

    S. Sukhbaatar, A. Szlam, J. Weston, R. Fergus, End-to-end memory networks, in: In Proc. 29th International Conference on Neural Information Processing Sys- tems, 2015, p. 2440–2448. 27

    2015

  34. [42]

    Liebel, M

    L. Liebel, M. K ¨orner, Auxiliary tasks in multi-task learning,https://arxiv. org/abs/1805.06334 (2018)

    2018 arXiv

  35. [43]

    Z. Yuan, W. Zhang, C. Tian, X. Rong, Z. Zhang, H. Wang, K. Fu, X. Sun, Remote sensing cross-modal text-image retrieval based on global and local information, IEEE Trans. Geosci. Remote Sens. 60 (2022) 1–16

    2022

  36. [44]

    J. Pan, Q. Ma, C. Bai, A prior instruction representation framework for remote sensing image-text retrieval, in: In Proc. 31st ACM International Conference on Multimedia, 2023, p. 611–620

    2023

  37. [45]

    S. Chen, C. Ge, Z. Tong, J. Wang, Y . Song, J. Wang, P. Luo, Adaptformer: Adapt- ing vision transformers for scalable visual recognition, in: In Proc. Advances in Neural Information Processing Systems, 2022, pp. 16664–16678

    2022

  38. [46]

    Jiang, J

    H. Jiang, J. Zhang, R. Huang, C. Ge, Z. Ni, J. Lu, J. Zhou, S. Song, G. Huang, Cross-modal adapter for text-video retrieval, https://arxiv.org/abs/ 2211.09623 (2022)

    2022 arXiv

  39. [47]

    H. Lu, Y . Huo, G. Yang, Z. Lu, W. Zhan, M. Tomizuka, M. Ding, Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling, in: In Proc. 12th International Conference on Learning Representations, 2024

    2024

  40. [48]

    K. Cha, D. Yu, J. Seo, Pushing the limits of vision-language models in remote sensing without human annotations, https://arxiv.org/abs/ 2409.07048 (2024). 28

    2024 arXiv

Pith tools

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