Finding Change in Satellite Archives from Text: How to Combine Before-and-After Images Efficiently

Giovanni Beltrame; Mark Bong; Simon Roy

REVIEW 2 major objections 6 minor 17 references

A cheap difference filter plus attention re-ranking finds text-described satellite changes 10–15× faster than full fusion while matching or beating its recall.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 03:25 UTC pith:PJVIVUKE

load-bearing objection Clean controlled bake-off with a real cascade win and a useful Mamba negative result; solid engineering, not a new paradigm. the 2 major comments →

arxiv 2607.28571 v1 pith:PJVIVUKE submitted 2026-07-30 cs.CV cs.IR

Finding Change in Satellite Archives from Text: How to Combine Before-and-After Images Efficiently

Simon Roy , Mark Bong , Giovanni Beltrame This is my paper

classification cs.CV cs.IR

keywords Earth observationchange retrievalvision-language modelsefficient architecturesstate space modelsbi-temporal fusioncascaded retrieval

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Operational archives need to answer queries like “find pairs where a new building appeared,” which means ranking many before-and-after satellite pairs by how well they match a text description of change. The expensive step at query time is the fusion module that combines the two views. Holding the image encoder and training recipe fixed, this paper compares eight fusion designs from attention, Mamba, and bottleneck-compression families on two benchmarks with ten seeds. Its main practical result is that a training-free cascade—rank the full gallery with simple subtraction, then re-rank only a short top-N list with attention fusion—matches or exceeds full-fusion recall on LEVIR-CC at 10–15× lower query cost, with quality parity on Dubai-CC. At ordinary vision-transformer patch counts, Mamba’s linear scan brings no wall-clock gain because it is memory-bound while attention parallelizes cleanly; bottleneck compression (TBF) cuts parameters 2.3× and latency 1.6× for a 0.007 change-only BLEU-1 cost. The work gives operators concrete, statistically grounded levers for cheaper text-based change search without inventing a new architecture.

Core claim

Under one frozen CLIP backbone and one training recipe, a training-free two-stage cascade—shortlist the full gallery with element-wise subtraction, then re-rank only the top-N with a trained attention fusion module—matches or exceeds full-fusion Recall@K on LEVIR-CC change queries while cutting per-query fusion cost 10–15×, and keeps R@1/R@5 parity on Dubai-CC. The same controlled eight-design comparison shows that Mamba variants yield no latency benefit at L=196, and that Temporal Bottleneck Fusion reduces parameters 2.3× and latency 1.6× for a 0.007 change-only BLEU-1 gap.

What carries the argument

Cascaded change retrieval: stage 1 ranks all candidate pairs with a cheap subtraction model; stage 2 re-ranks only the top-N shortlist with attention (or TBF) fusion, reusing already-trained models without retraining. Temporal Bottleneck Fusion (concat-then-reduce to width D before self-attention) is the companion compression pattern that attacks the doubled feature-width cost of stitching before and after views.

Load-bearing premise

That the retrieval-oriented LEVIR-CC split and strict instance-matching protocol used in the experiments are close enough to real archive search that the cascade’s recall preservation and the ranking of fusion families will transfer to operational settings.

What would settle it

On a larger, operationally split bi-temporal archive with diverse change types, check whether subtraction-then-attention at N=25 still matches or beats full-fusion R@1/R@5 while keeping the reported 10–15× latency cut; if cascade recall falls materially below full fusion, the central efficiency claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Operators can cut interactive query cost by an order of magnitude without retraining, by shortlisting with difference features before attention fusion.
At standard ViT tile sizes, optimized attention remains the default fusion choice; linear-time SSMs pay off only for longer multi-temporal stacks (L≳400).
TBF is a drop-in width-reduction pattern that roughly halves fusion latency and more than halves parameters for a negligible change-caption quality loss.
Fusion need not run on every archive pair at query time; cascade speedup grows toward the per-pair ceiling as gallery size increases.
Aggressive bottleneck widths should be checked on change-only metrics, because full-set aggregates can hide discarded change detail.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The cascade’s occasional recall gain over full fusion points to an implicit ensemble: candidates must score well under both a cheap difference signal and learned fusion, which may discard distractors either scorer alone would rank high.
The same shortlist-then-rerank pattern could extend to multi-temporal stacks beyond two frames, where full fusion cost grows faster.
Systems that already store CLIP embeddings offline could adopt the cascade with little more than a top-N handoff between a difference ranker and a fusion re-ranker.
The finding that memory-bound scans lose to parallel attention at L=196 may reappear in other short-sequence multimodal fusion tasks that currently default to linear-time SSMs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Clean controlled bake-off with a real cascade win and a useful Mamba negative result; solid engineering, not a new paradigm.

read the letter

The thing worth knowing is the cascade: subtraction shortlist then attention re-rank matches or beats full fusion on LEVIR-CC change queries at 10–15× lower query cost, with paired tests over ten seeds, and quality parity on Dubai-CC. Second is the wall-clock finding that selective-scan Mamba does not beat attention at L=196 on the same GPU—memory-bound scan vs parallel matmuls—with a length sweep that shows the crossover only past ~400 tokens.

What is actually new is the controlled eight-way comparison under one frozen encoder, one training recipe, and ten seeds, plus explicit cascade measurement and the Mamba latency negative. TBF itself is classical concat-then-reduce before self-attention; they do not oversell it. They ship code, report mean±std, Welch/paired tests, FLOPs and measured latency, a backbone check (GeoRSCLIP, SigLIP2), and a width note that aggressive compression hurts change-only metrics first. That is honest experimental craft.

Soft spots are ordinary, not load-bearing. LEVIR-CC uses a retrieval-oriented split enriched in change pairs, so absolute numbers do not line up with captioning papers and transfer to messy operational archives is unproven. Dubai-CC is tiny (300 train pairs) and noise-dominated. Latency is one desktop GPU. Strict instance Recall@K is low because many pairs share similar changes; they correctly add caption-similarity metrics and show distractors are often semantically fair. None of that undoes the within-benchmark cascade or efficiency ranking.

Math and citations look fine—no circular fitting, public CLIP/GeoRSCLIP/SigLIP2, fair pointers to Text-ITSR, ChangeMamba, BIT, bottleneck transformers. Free knobs (D=320, N, standard optim settings) are declared.

This is for people building text-based bi-temporal EO retrieval or efficient multimodal fusion who need deployable cost numbers, not a theory audience. I would send it to peer review; a serious editor should. Engage if you care about query-time fusion cost or SSM-vs-attention at ViT lengths; skip if you only want new architectures.

Referee Report

2 major / 6 minor

Summary. The paper studies the fusion module that combines frozen CLIP patch features from before-and-after satellite frames for text-based change retrieval. Under a single backbone, training recipe, and ten random seeds, it compares eight designs from attention, Mamba/SSM, and bottleneck-compression families (including Temporal Bottleneck Fusion, TBF) on LEVIR-CC and Dubai-CC. Three claims are advanced: (i) a training-free cascade that shortlists with element-wise subtraction then re-ranks with attention matches or exceeds full-fusion Recall@K on LEVIR-CC at 10–15× lower query cost (Table 4), with R@1/R@5 parity on Dubai-CC; (ii) Mamba’s linear scan yields no wall-clock gain at standard ViT length L=196 because the scan is memory-bound while attention maps to parallel matmuls, with a sequence-length sweep placing the crossover near L≈400; (iii) TBF cuts parameters ~2.3× and latency ~1.6× versus concatenation Transformers for a ~0.007 change-only BLEU-1 gap, while more aggressive width reduction is said to discard change-relevant detail.

Significance. If the cascade and efficiency rankings hold, the work gives concrete, immediately usable guidelines for operational text-based change search in EO archives: do not pay full fusion over the whole gallery; prefer bottlenecked attention over SSM fusion at tile-scale patch counts; and treat width compression as a first-order lever. Strengths that raise confidence include the controlled single-backbone/single-recipe design, mean±std over ten seeds with Welch and paired tests, measured latency/FLOPs on one GPU, a sequence-length sweep that explains the Mamba result, a backbone-robustness check (GeoRSCLIP, SigLIP2), and released code and split indices. The contribution is primarily empirical systems evidence rather than a new architecture, but the cascade result and the negative Mamba finding at L=196 are practically valuable and carefully reported.

major comments (2)

[§4.1 Bottleneck Compression; Abstract] Abstract and §4.1 state as a main finding that “more aggressive compression quietly discards change-relevant detail that aggregate metrics fail to reveal,” and that a width sweep shows b=320 matches the uncompressed width “within noise.” No table or figure reports that sweep (change-only vs full-set BLEU/Recall across bottleneck widths). Because this is listed among the three headline findings and is load-bearing for recommending TBF’s operating point over stronger compression, the width-ablation numbers (mean±std, change-only and full set) should be added—e.g., an appendix table—so the claim is inspectable rather than asserted in one sentence.
[§4.4 Cascaded Retrieval; Table 4] The cross-dataset cascade claim is asymmetric: Table 4 fully reports LEVIR-CC (N=25, paired per-seed tests, ms/query, speedup), while Dubai-CC is summarized in two sentences in §4.4 (“at N=50 the cascade matches full fusion on R@1/R@5 with about one point lower R@10, at 2.6× lower cost”) without a companion table or seed-level tests. Given that Dubai is already flagged as noise-dominated (§5) and is the only second benchmark, a small Dubai cascade table (R@K, cost, N) is needed to keep the “comparable R@1/R@5 on Dubai-CC” abstract claim on the same evidentiary footing as LEVIR.

minor comments (6)

[Fig. 1] Fig. 1 caption uses G=1929 and N=25 and quotes 907→89 ms; Table 1 single-pass latencies and Table 4 per-query costs are consistent in spirit but the caption could explicitly point to Table 4 so readers can reconcile single-pass vs gallery cost.
[§3.1] §3.1: state clearly whether text embeddings are also frozen CLIP text tower outputs and whether any text-side fine-tuning occurs; the visual side is explicit (“frozen CLIP”), the text side is only implied by InfoNCE on sij.
[Table 2] Table 2 caption says “scores average both retrieval directions”; a one-line definition of the reverse direction (image-pair→text vs text→image-pair) would remove ambiguity.
[§5 Limitations] §5 Limitations: latency on a single desktop GPU (RTX 4070) is acknowledged; a brief note on whether the cascade speedup ratio is expected to hold under batched GPU inference or CPU-only edge settings would help practitioners.
[Abstract; §1] Minor typos/spacing: “Weoutlinethreefindings”, “Changeretrieval”, “Vision–languagemod-els”, “andthe” style concatenations appear in the abstract/intro PDF text; clean for camera-ready.
[§2; §4.4] Related work could briefly cite classical IR cascade ranking beyond Wang et al. [15] only if space allows; not required for acceptance.

Circularity Check

0 steps flagged

No significant circularity: controlled empirical bake-off on external benchmarks with frozen public encoders; cascade reuses trained modules without refitting.

full rationale

The paper is a controlled architecture comparison, not a first-principles derivation. All eight fusion modules share one frozen CLIP (or GeoRSCLIP/SigLIP2) backbone, one InfoNCE recipe, and identical data splits; quality and latency are measured on held-out LEVIR-CC and Dubai-CC test galleries over ten seeds. The cascade claim is training-free reuse of already-trained Subtraction and attention modules (Table 4), not a parameter fit renamed as prediction. TBF is explicitly the classical concat-then-reduce pattern transferred to this task, not a renamed uniqueness theorem. References are to external prior work (CLIP, Mamba, Text-ITSR, BIT, etc.); there is no load-bearing self-citation chain and no equation that reduces a claimed prediction to its own fitted input by construction. Minor hyperparameter choices (D=320, N=25) are ordinary experimental settings, not circular steps. Honest finding: score 0, empty steps.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The work is an empirical systems comparison. It inherits standard contrastive VLM assumptions (frozen CLIP space is a meaningful similarity space for change, InfoNCE is a suitable objective) and standard benchmark constructions. No new physical entities are postulated; free parameters are ordinary architectural and optimization hyperparameters held fixed across the bake-off.

free parameters (3)

d_model / bottleneck width D=320 = 320
Chosen projection width before fusion; a width sweep is mentioned but the operating point is selected by the authors.
cascade shortlist size N = 25 (main table)
Top-N retained after subtraction prefilter; N=25 is reported as working across a range, but remains a design choice.
training hyperparameters (lr=8e-5, wd=5e-4, dropout=0.25, 30 epochs, batch 32, 3 layers) = as stated in Sect. 3.4
Single recipe applied to all modules; not swept per architecture.

axioms (4)

domain assumption Frozen CLIP ViT-B/16 patch features (L=196, Din=768) projected to D form a latent space in which cosine similarity after fusion ranks change descriptions correctly under InfoNCE.
Core modeling premise of Sect. 3.1; inherited from CLIP and Text-ITSR without re-derivation.
domain assumption Strict instance-level Recall@K plus n-gram overlap (BLEU/METEOR/ROUGE) against the five reference captions of the top-1 pair are adequate proxies for retrieval utility.
Stated in Sect. 3.1; the paper itself notes that strict matching underestimates practical utility when many pairs share similar changes.
domain assumption Wall-clock latency on one RTX 4070 with CUDA synchronization is representative of the relative efficiency ordering of attention vs selective-scan kernels.
Sect. 4.1 and 5; hardware-specific but standard for systems papers.
domain assumption The retrieval-oriented LEVIR-CC split and the Dubai-CC standard split are valid test beds for comparing fusion modules.
Sect. 3.2; authors flag that the LEVIR split differs from the captioning split.

invented entities (1)

Temporal Bottleneck Fusion (TBF) no independent evidence
purpose: Apply concat-then-reduce before self-attention so the Transformer operates at width D rather than 2D, cutting parameters and latency.
Named module in Sect. 3.3; it is a direct transfer of an established design pattern (FC-Siam-conc, BIT, Multimodal Bottleneck Transformer), not a new physical or mathematical entity.

pith-pipeline@v1.2.0-daily-grok45 · 14564 in / 3084 out tokens · 50759 ms · 2026-07-31T03:25:11.412319+00:00 · methodology

0 comments

read the original abstract

Operational Earth observation increasingly calls for answering queries such as ``find the image pairs where a new building appeared.'' This means searching an archive of before-and-after (bi-temporal) satellite image pairs and ranking each pair by how well it matches a natural-language description of the change. The component that performs this match, the fusion module that combines the ``before'' and ``after'' views, must be run at query time across many candidate pairs, so its speed largely sets the cost of every search. We present a controlled comparison of how to build that module. Using one fixed image encoder (a frozen CLIP model) and one training recipe for all variants, we evaluate eight designs drawn from three families: attention, state-space models (Mamba), and learned compression (our Temporal Bottleneck Fusion, TBF). Each design is tested on two benchmarks (LEVIR-CC and Dubai-CC) with ten random seeds, so the reported differences are statistically grounded. We outline three findings: first, a training-free two-stage search (a cheap difference model that shortlists candidates, followed by attention fusion that re-ranks them) matches or exceeds full-fusion recall on LEVIR-CC while cutting query cost $10$-$15\times$, with comparable R@1/R@5 on Dubai-CC; second, the linear-time scan of Mamba, attractive on paper, gives no speed benefit at the patch counts typical of vision transformers ($L{=}196$): the scan is limited by memory bandwidth, whereas attention maps cleanly onto parallel hardware; and third, compressing the fused representation (TBF) reduces parameters by $2.3\times$ and latency by $1.6\times$ for a change-only BLEU-1 cost of $0.007$, although more aggressive compression quietly discards change-relevant detail that aggregate metrics fail to reveal.

Figures

Figures reproduced from arXiv: 2607.28571 by Giovanni Beltrame, Mark Bong, Simon Roy.

**Figure 1.** Figure 1: Cascaded change retrieval (LEVIR-CC, G=1929, N=25). Frames are CLIPencoded offline; a subtraction filter ranks all candidate pairs, then TBF re-ranks the top-N shortlist, cutting fusion latency from 907 to 89 ms per query while matching or exceeding full-fusion recall. image. Assessing storm damage or auditing urban expansion requires comparing two acquisitions, e.g., finding every tile where “a building … view at source ↗

**Figure 2.** Figure 2: (a) Recall@1 vs. fusion latency on LEVIR-CC change queries (CT = Concatenation Transformer; dashed = Pareto front). (b) TBF top-3 retrievals; the correct pair (green) ranks first (top) and third (bottom), and the distractors (brown) are semantically plausible [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 7 linked inside Pith

[1]

IEEE Transactions on Geoscience and Remote Sensing60, 1–14 (2022)

Chen, H., Qi, Z., Shi, Z.: Remote sensing image change detection with trans- formers. IEEE Transactions on Geoscience and Remote Sensing60, 1–14 (2022). https://doi.org/10.1109/TGRS.2021.3095166

arXiv 2022
[2]

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

Chen, H., Song, J., Han, C., Xia, J., Yokoya, N.: Changemamba: Remote sensing change detection with spatiotemporal state space model. IEEE Transactions on Geoscience and Remote Sensing62, 1–20 (2024)

2024
[3]

In: International Conference on Machine Learning (ICML) (2024)

Dao, T., Gu, A.: Transformers are SSMs: Generalized models and efficient al- gorithms through structured state space duality. In: International Conference on Machine Learning (ICML) (2024)

2024
[4]

In: 2018 25th IEEE International Conference on Image Process- ing (ICIP)

Daudt, R.C., Le Saux, B., Boulch, A.: Fully convolutional siamese networks for change detection. In: 2018 25th IEEE International Conference on Image Process- ing (ICIP). pp. 4063–4067. IEEE (2018)

2018
[5]

Ferrod, R., Caro, L.D., Ienco, D.: Towards a multimodal framework for remote sensing image change retrieval and captioning (2024), arXiv:2406.13424

Pith/arXiv arXiv 2024
[6]

Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces (2024), arXiv:2312.00752

Pith/arXiv arXiv 2024
[7]

Hoxha, G., Angyal, O., Demir, B.: Self-supervised cross-modal text-image time series retrieval in remote sensing (2025), arXiv:2501.19043

Pith/arXiv arXiv 2025
[8]

IEEE Transactions on Geoscience and Remote Sensing60, 1–14 (2022)

Hoxha, G., Chouaf, S., Melgani, F., Smara, Y.: Change captioning: A new paradigm for multitemporal remote sensing image analysis. IEEE Transactions on Geoscience and Remote Sensing60, 1–14 (2022). https://doi.org/10.1109/TGRS.2022.3195692

arXiv 2022
[9]

IEEE Transactions on Geoscience and Remote Sensing60, 1–20 (2022)

Liu, C., Zhao, R., Chen, H., Zou, Z., Shi, Z.: Remote sensing image change captioning with dual-branch transformers: A new method and a large scale dataset. IEEE Transactions on Geoscience and Remote Sensing60, 1–20 (2022). https://doi.org/10.1109/TGRS.2022.3218921

arXiv 2022
[10]

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 (2024), arXiv:2306.11029

Pith/arXiv arXiv 2024
[11]

Nagrani, A., Yang, S., Arnab, A., Jansen, A., Schmid, C., Sun, C.: Attention bottlenecks for multimodal fusion (2022), arXiv:2107.00135

Pith/arXiv arXiv 2022
[12]

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021)

2021
[13]

Sharma, S., El Asri, L., Schulz, H., Zumer, J.: Relevance of unsupervised met- rics in task-oriented dialogue for evaluating natural language generation (2017), arXiv:1706.09799

Pith/arXiv arXiv 2017
[14]

Tschannen, M., Gritsenko, A., Wang, X., Naeem, M.F., Alabdulmohsin, I., et al.: SigLIP 2: Multilingual vision-language encoders with improved semantic under- standing, localization, and dense features (2025)

2025
[15]

In: Proc

Wang, L., Lin, J., Metzler, D.: A cascade ranking model for efficient ranked re- trieval. In: Proc. ACM SIGIR. pp. 105–114 (2011)

2011
[16]

IEEE Transactions on Geoscience and Remote Sensing62, 1–23 (2024)

Zhang, Z., Zhao, T., Guo, Y., Yin, J.: Rs5m and GeoRSCLIP: A large-scale vision- language dataset and a large vision-language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing62, 1–23 (2024)

2024
[17]

Zhao, H., Zhang, M., Zhao, W., Ding, P., Huang, S., Wang, D.: Cobra: Extend- ing mamba to multi-modal large language model for efficient inference (2025), arXiv:2403.14520

Pith/arXiv arXiv 2025

[1] [1]

IEEE Transactions on Geoscience and Remote Sensing60, 1–14 (2022)

Chen, H., Qi, Z., Shi, Z.: Remote sensing image change detection with trans- formers. IEEE Transactions on Geoscience and Remote Sensing60, 1–14 (2022). https://doi.org/10.1109/TGRS.2021.3095166

arXiv 2022

[2] [2]

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

Chen, H., Song, J., Han, C., Xia, J., Yokoya, N.: Changemamba: Remote sensing change detection with spatiotemporal state space model. IEEE Transactions on Geoscience and Remote Sensing62, 1–20 (2024)

2024

[3] [3]

In: International Conference on Machine Learning (ICML) (2024)

Dao, T., Gu, A.: Transformers are SSMs: Generalized models and efficient al- gorithms through structured state space duality. In: International Conference on Machine Learning (ICML) (2024)

2024

[4] [4]

In: 2018 25th IEEE International Conference on Image Process- ing (ICIP)

Daudt, R.C., Le Saux, B., Boulch, A.: Fully convolutional siamese networks for change detection. In: 2018 25th IEEE International Conference on Image Process- ing (ICIP). pp. 4063–4067. IEEE (2018)

2018

[5] [5]

Ferrod, R., Caro, L.D., Ienco, D.: Towards a multimodal framework for remote sensing image change retrieval and captioning (2024), arXiv:2406.13424

Pith/arXiv arXiv 2024

[6] [6]

Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces (2024), arXiv:2312.00752

Pith/arXiv arXiv 2024

[7] [7]

Hoxha, G., Angyal, O., Demir, B.: Self-supervised cross-modal text-image time series retrieval in remote sensing (2025), arXiv:2501.19043

Pith/arXiv arXiv 2025

[8] [8]

IEEE Transactions on Geoscience and Remote Sensing60, 1–14 (2022)

Hoxha, G., Chouaf, S., Melgani, F., Smara, Y.: Change captioning: A new paradigm for multitemporal remote sensing image analysis. IEEE Transactions on Geoscience and Remote Sensing60, 1–14 (2022). https://doi.org/10.1109/TGRS.2022.3195692

arXiv 2022

[9] [9]

IEEE Transactions on Geoscience and Remote Sensing60, 1–20 (2022)

Liu, C., Zhao, R., Chen, H., Zou, Z., Shi, Z.: Remote sensing image change captioning with dual-branch transformers: A new method and a large scale dataset. IEEE Transactions on Geoscience and Remote Sensing60, 1–20 (2022). https://doi.org/10.1109/TGRS.2022.3218921

arXiv 2022

[10] [10]

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 (2024), arXiv:2306.11029

Pith/arXiv arXiv 2024

[11] [11]

Nagrani, A., Yang, S., Arnab, A., Jansen, A., Schmid, C., Sun, C.: Attention bottlenecks for multimodal fusion (2022), arXiv:2107.00135

Pith/arXiv arXiv 2022

[12] [12]

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021)

2021

[13] [13]

Sharma, S., El Asri, L., Schulz, H., Zumer, J.: Relevance of unsupervised met- rics in task-oriented dialogue for evaluating natural language generation (2017), arXiv:1706.09799

Pith/arXiv arXiv 2017

[14] [14]

Tschannen, M., Gritsenko, A., Wang, X., Naeem, M.F., Alabdulmohsin, I., et al.: SigLIP 2: Multilingual vision-language encoders with improved semantic under- standing, localization, and dense features (2025)

2025

[15] [15]

In: Proc

Wang, L., Lin, J., Metzler, D.: A cascade ranking model for efficient ranked re- trieval. In: Proc. ACM SIGIR. pp. 105–114 (2011)

2011

[16] [16]

IEEE Transactions on Geoscience and Remote Sensing62, 1–23 (2024)

Zhang, Z., Zhao, T., Guo, Y., Yin, J.: Rs5m and GeoRSCLIP: A large-scale vision- language dataset and a large vision-language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing62, 1–23 (2024)

2024

[17] [17]

Zhao, H., Zhang, M., Zhao, W., Ding, P., Huang, S., Wang, D.: Cobra: Extend- ing mamba to multi-modal large language model for efficient inference (2025), arXiv:2403.14520

Pith/arXiv arXiv 2025