Leveraging Language Prior for Infrared Small Target Detection

Pranav Singh; Pravendra Singh

REVIEW 4 major objections 4 minor 15 references

Leveraging Language Prior for Infrared Small Target Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Textual descriptions of target location, generated once by a vision-language model and used only as a training signal, improve infrared small-target detection across two public benchmarks, lifting IoU on IRSTD-1k from 69.99 to 73.30.

desk verdict A genuinely new application of language-guided attention to IRSTD with a useful dataset; the core effect is consistent, but the mechanism is unverified and a table inconsistency muddies the numbers. read the letter →

arxiv 2507.13113 v1 pith:VCPORWLW submitted 2025-07-17 cs.CV

Pranav Singh , Pravendra Singh This is my paper

classification cs.CV

keywords infraredsmalltargetdetectionlanguagepriormultimodallearningvision-languagemodelsCLIPattentionguidanceLangIRdatasetsegmentation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that adding short natural-language descriptions of where a small infrared target sits (e.g., 'center', 'upper left') as a training signal makes a U-shaped segmentation network detect those targets more accurately. The authors introduce LGNet, a UNet-style encoder-decoder that fuses a CLIP-derived text embedding of such a description with the image embedding to create a target descriptor, which then rescales feature maps via learned attention weights. They build their LangIR dataset by asking the GPT-4 vision model to write 50-word location descriptions for images in the IRSTD-1k and NUAA-SIRST datasets, and report consistent gains over image-only baselines, for example raising IoU on IRSTD-1k from 69.99 to 73.30 with language used only during training. If the claim holds, it gives IRSTD practitioners a cheap way to boost detection without changing the test-time pipeline.

What carries the argument

The load-bearing components are (1) the Target Descriptor, obtained by element-wise addition of a CLIP text embedding of the generated description and a CLIP image embedding of the infrared frame; and (2) the language fusion block, which takes the global-average-pooled encoder feature vector $v_1$ and an MLP-mapped version $v_2$ of the Target Descriptor, alternately stacks them, then applies group-wise convolution, batch normalization, and sigmoid to compute feature-wise attention weights that gate the last encoder and decoder outputs. The same attention mechanism, without the language stream, forms the ordinary fusion blocks throughout the network. Textual descriptions are produced in advance by the GPT-4 vision model using a system prompt that constrains the answer to about 50 words, and the paper's ablation shows that the language-guided attention is what converts an image-only baseline (IoU 69.99) into the reported 73.30.

What would settle it

Take the same LGNet training pipeline but replace each real description with a randomly shuffled collection of the same location words (for instance, swap 'center' with 'upper left' between images). If IoU stays near 73.30, the semantic content of the language prior is not what drives the improvement. Alternatively, run a CLIP retrieval test on LangIR: given a description, rank all infrared images by CLIP image-embedding similarity; if the correct image is not ranked at or near the top, the presumed text-image alignment underlying the Target Descriptor is not present.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a coarse linguistic location prior — a phrase like 'the small target, a boat, is located in the center of the image, appearing as a bright spot on the contrasting darker water surface' — can be turned, through CLIP text embeddings, into an effective attention guide for infrared small target detection. The target descriptor formed by element-wise addition of the CLIP text and image embeddings is passed through a lightweight fusion block that produces sigmoid attention weights, which modulate the deepest encoder and decoder features of LGNet. Applying this language guidance only during training lifts IoU on IRSTD-1k from 69.99 (no language) to 73.30, while remaining at 73.35 when the description is also given at test time, and similar gains appear on the NUAA-SIRST subset. The authors interpret this as evidence that text provides a beneficial semantic supervision signal for learning discriminative features, even when text is absent at inference.

Load-bearing premise

The results hinge on the assumption that CLIP text embeddings for coarse location words like 'left' or 'center' align meaningfully with CLIP image embeddings of infrared small targets, even though CLIP was trained on natural images; if that alignment is weak, the element-wise addition in the Target Descriptor would put noise into the attention weights and the reported gains would need another explanation.

Editorial extensions

If this is right

Because the language prior is used only in training, adopting LGNet adds no textual dependency at run time; a deployed system needs only the image encoder and the same 25 ms per-frame inference.
The pattern of gains on both IRSTD-1k and NUAA-SIRST suggests the result transfers across infrared datasets with different target sizes and backgrounds, and across text sources (GPT-4V vs Claude 3.5 Sonnet), implying the prior is what matters rather than the specific VLM.
Training with language raises detection probability (Pd to 98.26 on IRSTD-1k) while cutting false alarms (Fa to 4.09), so the benefit is not just overlap but cleaner target decisions.
The approach supplies the first image-text paired dataset (LangIR, 1,428 pairs) for IRSTD, which future multimodal training and evaluation can build on.

Reading between the lines

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

Because the paper's quadrant-only ablation still gives IoU 72.4 versus 73.30 with full descriptions, much of the gain may come from coarse region information rather than richer semantics; a direct test would be training with one-hot quadrant vectors instead of CLIP text embeddings to see whether the margin over the image-only baseline is preserved.
The same training-only language guidance could plausibly transfer to other small-object detection domains (medical, aerial) where CLIP text-image alignment is valid; the paper does not test this, and per-domain alignment should be checked before assuming it.
The paper leaves open whether the language prior works by regularizing the network or by providing a soft location label; comparing LGNet against an image-only network with equivalent added capacity or dropout would isolate the mechanism.

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.

Desk Editor's Note

read the letter

The paper does something new: it applies language-guided attention to infrared small target detection and contributes a VLM-generated text dataset (LangIR) built on existing image datasets. The practical choice of using language only during training, with no test-time overhead, is sensible, and the ablations are more thorough than what we usually see in this subfield.

The main claim, that language supervision during training improves IRSTD by a few IoU points, is consistent across the ablations. The quadrant-only versus full-semantic-text comparison (72.4 vs. 73.3 IoU) is a nice hint that richer text matters, and the prompt-style study shows real care. The LangIR dataset could be a useful resource, assuming the authors actually release it.

Now the soft spots, in proportion. First, Tables 2 and 6 report different nIoU and Fa for what should be the same baseline (67.79 vs. 67.89; 6.43 vs. 4.09). That is not a harmless typo. It changes the claimed Fa improvement and needs to be reconciled. Second, the "first multimodal IRSTD" claim sits awkwardly next to SAIST, which appears in their own comparison table and also uses contrastive language-image pretraining; that overstatement should be fixed. Third, and most important, the mechanism is not established. The paper adds CLIP text and image embeddings element-wise, but it never shows that these embeddings actually align on infrared imagery. The "fine-tuned CLIP text encoder" is mentioned and never described. The improvement could come from a generic perturbation effect rather than semantic language guidance. A control that shuffles text descriptions across training images would separate semantics from noise, and the paper does not include it. That is not a fatal flaw, but it leaves the central "language prior" claim unproven. Minor points: no code or dataset link is given, and no error bars or multiple seeds are reported.

This is a conditional-accept caliber paper, not a desk reject. It deserves a serious referee, and the referee should push on the table inconsistency, the randomized-text control, and the "first multimodal" wording. For anyone working on IRSTD or vision-language detection, it is worth reading and likely worth citing for the dataset once it is public.

Referee Report

4 major / 4 minor

Summary. The paper proposes LGNet, a U-Net-style detector for infrared small target detection (IRSTD) that fuses a so-called language prior with image features. Text descriptions of target locations are generated by GPT-4V from each infrared image, encoded by a CLIP text encoder, and added element-wise to CLIP image embeddings to form a target descriptor. A language-guided attention block modulates the last encoder and decoder feature maps. The language is used only during training; at inference only image embeddings are used. The authors introduce the LangIR dataset, adding text descriptions to IRSTD-1k and NUAA-SIRST, and report improved IoU, nIoU, Pd, and Fa over prior single-modality methods, together with ablations on prompt style, VLM choice, and component contributions.

Significance. If the reported gains are real and attributable to the language content, this is a useful and timely direction: it shows that off-the-shelf vision-language models can provide auxiliary supervision for a task with tiny, low-contrast targets where manual annotation is difficult, and the released LangIR dataset would be a resource for follow-up work. The paper also deserves credit for including model-agnostic VLM generation (GPT-4V vs. Claude), a practical training-only language setting, and several focused ablations. However, the central claim as stated is not yet established: the text is generated from the same image being detected, so it is a re-encoding of image content rather than an independent external prior, and the paper provides no control that isolates the semantic contribution of the text from a generic perturbation effect. The missing details of the fine-tuned CLIP text encoder and the internal inconsistencies in the ablation tables further weaken the current support for the mechanism.

major comments (4)

[Table 2 vs. Table 6] The baseline configuration is reported inconsistently across the two tables. In Table 2, the 'Never (Baseline)' row gives nIoU = 67.79 and Fa = 6.43, while in Table 6 the 'Baseline' row gives nIoU = 67.89 and Fa = 4.09 for what should be the same network and data split. In fact, the Fa value 4.09 in Table 6 'Baseline' matches the 'Training Only' row of Table 2, not the baseline row. Since Table 6 is the central evidence for the contribution of the language fusion block, these numbers must be corrected and re-verified, and the source of the discrepancy explained.
[Sec. 3.2, Sec. 5.2] The 'fine-tuned CLIP text encoder' is mentioned in Sec. 3.2 but never described. No fine-tuning dataset, objective, training protocol, or hyperparameters are provided, and it is not stated whether the CLIP image encoder is also fine-tuned. This missing detail makes the target descriptor generation irreproducible. It also matters for the scientific claim, because the element-wise addition Te + Ie assumes that the text and image embeddings are aligned for infrared inputs, and CLIP was trained on natural images rather than infrared imagery. The authors should either specify the fine-tuning procedure or remove the claim of fine-tuning.
[Sec. 5.6.4 and Eqs. (1)-(3)] The experiments do not include a mismatched-text control. Because the GPT-4V descriptions are generated from the very image being detected, the language input is not independent of the image; it is a self-generated auxiliary signal. The improvement from 'Training Only' over 'Never' could therefore arise from a generic per-sample perturbation of v2 in the language fusion block rather than from the semantic content of the text. The quadrant-only ablation in Sec. 5.6.4 varies the richness of the text but keeps the text-image correspondence intact. A control that randomly permutes descriptions across training images, and ideally also a fixed-description control, is necessary to attribute the observed gain to language semantics. Without such a control, the central 'language prior' claim is not established.
[Sec. 5.1 and Tables 2, 4, 5] No error bars, multiple runs, or statistical significance tests are reported. Given that the difference between 'Training Only' (IoU 73.30) and 'Training+Test' (IoU 73.35) is only 0.05 IoU, single-run differences of this size cannot support the claim of 'significant enhancement.' The main comparisons in Tables 4 and 5 and the ablation in Table 6 should be repeated over at least three seeds and reported as mean ± std, so the reader can judge whether the observed improvements are stable.

minor comments (4)

[Table 6] The header 'Varient' is a typo for 'Variant.'
[Abstract and Sec. 5.4] The abstract reports 'relative percentage differences' of 67.87% and 113.43% for Fa. Since lower Fa is better, the sign of the improvement is ambiguous; stating that Fa is reduced by these percentages would be clearer.
[Sec. 5.3.1] The comparison of inference times states LGNet requires 37 ms (25 + 11.6 ms) versus TCI-Former's 44 ms, but this excludes the VLM text generation time that would be needed in the 'Training+Test' setting. Please clarify that the 25 ms figure applies to the image-only inference mode.
[References] Several references contain garbled author lists (e.g., Zhang et al., 2025a and Zhang et al., 2025b), and the method name is spelled inconsistently as both 'RepsISD' and 'RepISD' in Table 4. The reference list should be cleaned up.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported language-prior gain rests on empirical training-time ablations, not on a derivation that reduces to the method's inputs.

full rationale

The paper's central claim is that adding a text-derived target descriptor during training improves infrared small-target detection while keeping inference unchanged. This is established by direct empirical comparisons (Table 2: Never vs. Training Only; Table 6: +Fusion Block vs. +Language Fusion Block), which are legitimate contrasts. The language prior is indeed generated from the same infrared image by GPT-4V (Sec. 3.1), so it is not an external prior in the information-theoretic sense, and the absence of a shuffled-text control leaves the semantic-alignment mechanism unverified. Those are validity and interpretation concerns, not circularity: no equation defines the reported IoU in terms of the language prior, no fitted parameter is renamed as a prediction, and no load-bearing claim rests on a self-citation. The under-specified 'fine-tuned CLIP text encoder' is a missing-detail issue rather than a circular step. Accordingly, the derivation chain is self-contained as an empirical study, and no circular step can be exhibited with the required specificity.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The central claim does not rest on fitted constants or invented physical entities. The model parameters are learned from data via BCE loss, which is standard for deep learning and not a free parameter in the derivation sense. The main conceptual postulates are the two domain assumptions above regarding CLIP alignment and VLM reliability.

assumptions (2)

domain assumption CLIP's text and image embeddings remain meaningfully aligned for infrared images despite being pre-trained on natural images.
Section 3.2 constructs the target descriptor as Te + Ie and relies on this alignment for the language-guided attention. No infrared-specific fine-tuning of CLIP is demonstrated, and the paper's wording on fine-tuning is inconsistent.
domain assumption The GPT-4 Vision model produces text descriptions that locate small targets accurately enough to serve as a training signal.
Section 5.3 describes prompt engineering to elicit locations, but the VLM is shown to hallucinate when the word "targets" is used (Sec. 5.6.6) and to fail at direct bounding box annotation (Sec. 5.6.2). The accuracy of the generated text is not quantitatively evaluated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Leveraging Language Prior for Infrared Small Target Detection." pith.science (2026). https://pith.science/paper/VCPORWLW

@misc{pith2026250713113,
  author       = {Pith},
  title        = {Pith review of: Leveraging Language Prior for Infrared Small Target Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VCPORWLW}},
  note         = {Machine review of arXiv:2507.13113}
}

read the original abstract

IRSTD (InfraRed Small Target Detection) detects small targets in infrared blurry backgrounds and is essential for various applications. The detection task is challenging due to the small size of the targets and their sparse distribution in infrared small target datasets. Although existing IRSTD methods and datasets have led to significant advancements, they are limited by their reliance solely on the image modality. Recent advances in deep learning and large vision-language models have shown remarkable performance in various visual recognition tasks. In this work, we propose a novel multimodal IRSTD framework that incorporates language priors to guide small target detection. We leverage language-guided attention weights derived from the language prior to enhance the model's ability for IRSTD, presenting a novel approach that combines textual information with image data to improve IRSTD capabilities. Utilizing the state-of-the-art GPT-4 vision model, we generate text descriptions that provide the locations of small targets in infrared images, employing careful prompt engineering to ensure improved accuracy. Due to the absence of multimodal IR datasets, existing IRSTD methods rely solely on image data. To address this shortcoming, we have curated a multimodal infrared dataset that includes both image and text modalities for small target detection, expanding upon the popular IRSTD-1k and NUDT-SIRST datasets. We validate the effectiveness of our approach through extensive experiments and comprehensive ablation studies. The results demonstrate significant improvements over the state-of-the-art method, with relative percentage differences of 9.74%, 13.02%, 1.25%, and 67.87% in IoU, nIoU, Pd, and Fa on the NUAA-SIRST subset, and 4.41%, 2.04%, 2.01%, and 113.43% on the IRSTD-1k subset of the LangIR dataset, respectively.

Figures

Figures reproduced from arXiv: 2507.13113 by the authors.

**Figure 1.** Illustration of the language prior used for IRSTD task. The text and image embeddings from CLIP (Radford et al., 2021) are combined element-wise to produce the Target Descriptor, which is then incorporated into our proposed LGNet for small target detection. from multiple modalities can significantly enhance the effectiveness of the small target detection task compared to relying on inputs from a single modality. Our… view at source ↗

**Figure 2.** Illustration of the system prompt template utilized for generating textual descriptions using the GPT-4 vision model (VLM). The input image is parsed into a base64 string format and then inserted into the input field of the prompt template. We use VLM in the system role, followed by the task description for the VLM. The prompt is then given to the GPT-4 Vision model to generate a textual description that accurately … view at source ↗

**Figure 3.** The overview of our proposed LGNet. LGNet is based on UNet architecture with encoder-decoder blocks. Each encoder-decoder is a residual U-Block (shown on the right). The feature maps are encoded gradually, and at the same time, the decoder decodes the feature maps. The outputs of the adjacent encoder-decoders are fused in the fusion block and then passed on to the next decoder. Furthermore, the language prior are us… view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: Illustration of the language fusion block (left) and the fusion block (right) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Illustration of the alternative stacking in the language fusion block, where 𝑣1 and 𝑣2 are combined alternately. This combination then undergoes group-wise convolution, batch normalization, and a sigmoid function to produce the language-guided attention weights. the in…

**Figure 6.** Figure 6: Experimental results with three different prompt styles to guide small target detection. ‘Sys’ represents the system prompt, ‘Few’ represents the few-shot prompt, and ‘Zero’ represents the zero-shot prompt. The corresponding ground truth refers to the segmentation mask…

**Figure 7.** Figure 7: Qualitative results obtained using LGNet on the IRSTD-1K dataset. Circles indicate small targets, with an enlarged version shown in the box. The illustration shows LGNet’s robustness across varying IRSTD conditions. Row 1 shows detection in a complex background with cl…

**Figure 8.** Figure 8: Illustration of direct small target detection by the VLM. The VLM is not able to accurately annotate the target because it draws a box that does not include the small target. However, the generated text description is correct, as the VLM uses the term ‘lower right quad…

**Figure 9.** Figure 9: Effect of word limit constraint on responses: the response in green is generated with a 50-word limit, while the response in purple shows the VLM-generated description when no word limit constraint is applied in the prompt. noted that the VLM can detect the object’s lo…

**Figure 10.** Figure 10: The VLM can detect multiple targets when using the word ‘target’ in the system prompt template, as shown, where it identifies multiple small targets in the response (shown in green color box) for the IRSTD-1k dataset. The VLM hallucinates when the word ‘targets’ is us…

**Figure 11.** Figure 11: Dataset format of NUAA-SIRST and IRSTD-1k subsets of LangIR. IR images are in the images directory, with the corresponding detection masks in the masks directory. The text descriptions of the small targets are in the description directory [PITH_FULL_IMAGE:figures/ful…

**Figure 12.** Figure 12: Illustration of the textual descriptions for small target detection in infrared images. (Top) The textual description of the targets in the IRSTD-1k dataset indicates the target object’s location in the text. (Bottom) The textual description of the targets in the NUAA…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 9 canonical work pages

[1]

Gpt-4 technical report

Achiam,J.,Adler,S.,Agarwal,S.,Ahmad,L.,Akkaya,I.,Aleman,F.L.,Almeida,D.,Altenschmidt,J.,Altman,S.,Anadkat,S.,etal.,2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 . Bao, C., Cao, J., Ning, Y., Zhao, T., Li, Z., Wang, Z., Zhang, L., Hao, Q.,

arXiv 2023
[6]

IEEE Transactions on Geoscience and Remote Sensing

Diffusion-based continuous feature representation for infrared small-dim target detection. IEEE Transactions on Geoscience and Remote Sensing . Gao,C.,Meng,D.,Yang,Y.,Wang,Y.,Zhou,X.,Hauptmann,A.G.,2013. Infraredpatch-imagemodelforsmalltargetdetectioninasingleimage. IEEE transactions on image processing 22, 4996–5009. Han,J.,Moradi,S.,Faramarzi,I.,Liu,C.,...

work page 2013
[8]

Pattern recognition 106, 107404

U2-net: Going deeper with nested u-structure for salient object detection. Pattern recognition 106, 107404. Radford,A.,Kim,J.W.,Hallacy,C.,Ramesh,A.,Goh,G.,Agarwal,S.,Sastry,G.,Askell,A.,Mishkin,P.,Clark,J.,etal.,2021.Learningtransferable visual models from natural language supervision, in: International conference on machine learning, PMLR. pp. 8748–8763...

work page 2021
[9]

IET image processing 14, 1937–1947

Reweighted infrared patch image model for small target detection based on non-convex lp-norm minimisation and tv regularisation. IET image processing 14, 1937–1947. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.,

work page 1937
[10]

arXiv preprint arXiv:2401.14159

Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 . Shangguan, Z., Seita, D., Rostami, M.,

arXiv
[11]

Cross-domain Multi-modal Few-shot Object Detection via Rich Text

Cross-domain multi-modal few-shot object detection via rich text. arXiv preprint arXiv:2403.16188 . Shen, R., Inoue, N., Shinoda, K.,

work page Pith review arXiv
[12]

arXiv preprint arXiv:2302.13971

Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 . Wang, H., Zhou, L., Wang, L.,

arXiv
[13]

arXiv preprint arXiv:2311.11829

System 2 attention (is something you might need too). arXiv preprint arXiv:2311.11829 . Wu, F., Zhang, T., Li, L., Huang, Y., Peng, Z.,

arXiv

Show all 15 references

[382]

Rkformer:Runge-kuttatransformerwithrandom-connectionattention for infrared small target detection, in: Proceedings of the 30th ACM International Conference on Multimedia, pp

Zhang,M.,Bai,H.,Zhang,J.,Zhang,R.,Wang,C.,Guo,J.,Gao,X.,2022a. Rkformer:Runge-kuttatransformerwithrandom-connectionattention for infrared small target detection, in: Proceedings of the 30th ACM International Conference on Multimedia, pp. 1730–1738. Zhang, M., Li, Xiaolong, F.,...

2024
[2012]

Contextual object detection using set-based classification, in: Computer Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, Springer. pp. 43–57. Dai, Y., Wu, Y.,

2012
[2019]

arXiv preprint arXiv:1910.13461

Bart: Denoising sequence-to- sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461 . Li,B.,Wang,Y.,Wang,L.,Zhang,F.,Liu,T.,Lin,Z.,An,W.,Guo,Y.,2023. Montecarlolinearclusteringwithsingle-pointsupervisionisenough f...

1910 arXiv
[2020]

Advances in neural information processing systems 33, 1877–1901

Language models are few-shot learners. Advances in neural information processing systems 33, 1877–1901. Chen, C.P., Li, H., Wei, Y., Xia, T., Tang, Y.Y.,

1901
[2022]

arXiv preprint arXiv:2208.06677

Adan: Adaptive nesterov momentum algorithm for faster optimizing deep models. arXiv preprint arXiv:2208.06677 . Zhang, L., Peng, Z.,

arXiv
[2023]

arXiv preprint arXiv:2311.08747

Improved dense nested attention network based on transformer for infrared small target detection. arXiv preprint arXiv:2311.08747 . Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan,J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.,

arXiv
[2024]

arXiv preprint arXiv:2402.02046

Tci-former: Thermal conduction-inspired transformer for infrared small target detection. arXiv preprint arXiv:2402.02046 . Cinbis, R.G., Sclaroff, S.,

arXiv

Pith tools

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

[1] [1]

Gpt-4 technical report

Achiam,J.,Adler,S.,Agarwal,S.,Ahmad,L.,Akkaya,I.,Aleman,F.L.,Almeida,D.,Altenschmidt,J.,Altman,S.,Anadkat,S.,etal.,2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 . Bao, C., Cao, J., Ning, Y., Zhao, T., Li, Z., Wang, Z., Zhang, L., Hao, Q.,

arXiv 2023

[2] [6]

IEEE Transactions on Geoscience and Remote Sensing

Diffusion-based continuous feature representation for infrared small-dim target detection. IEEE Transactions on Geoscience and Remote Sensing . Gao,C.,Meng,D.,Yang,Y.,Wang,Y.,Zhou,X.,Hauptmann,A.G.,2013. Infraredpatch-imagemodelforsmalltargetdetectioninasingleimage. IEEE transactions on image processing 22, 4996–5009. Han,J.,Moradi,S.,Faramarzi,I.,Liu,C.,...

work page 2013

[3] [8]

Pattern recognition 106, 107404

U2-net: Going deeper with nested u-structure for salient object detection. Pattern recognition 106, 107404. Radford,A.,Kim,J.W.,Hallacy,C.,Ramesh,A.,Goh,G.,Agarwal,S.,Sastry,G.,Askell,A.,Mishkin,P.,Clark,J.,etal.,2021.Learningtransferable visual models from natural language supervision, in: International conference on machine learning, PMLR. pp. 8748–8763...

work page 2021

[4] [9]

IET image processing 14, 1937–1947

Reweighted infrared patch image model for small target detection based on non-convex lp-norm minimisation and tv regularisation. IET image processing 14, 1937–1947. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.,

work page 1937

[5] [10]

arXiv preprint arXiv:2401.14159

Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 . Shangguan, Z., Seita, D., Rostami, M.,

arXiv

[6] [11]

Cross-domain Multi-modal Few-shot Object Detection via Rich Text

Cross-domain multi-modal few-shot object detection via rich text. arXiv preprint arXiv:2403.16188 . Shen, R., Inoue, N., Shinoda, K.,

work page Pith review arXiv

[7] [12]

arXiv preprint arXiv:2302.13971

Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 . Wang, H., Zhou, L., Wang, L.,

arXiv

[8] [13]

arXiv preprint arXiv:2311.11829

System 2 attention (is something you might need too). arXiv preprint arXiv:2311.11829 . Wu, F., Zhang, T., Li, L., Huang, Y., Peng, Z.,

arXiv