Prompt-Driven Simulation with Feature Perturbation for Cross-Domain Few-Shot Object Detection

Junxi Cai; Linhai Zhuo; Qingping Zheng; Tianwen Qian; Yang Liu

REVIEW 4 major objections 6 minor 1 cited by

Prompt-Driven Simulation with Feature Perturbation for Cross-Domain Few-Shot Object Detection

T0 review · 4 major / 6 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 Prompt-driven VLM generation plus feature perturbation improves cross-domain few-shot object detection on six target domains.

desk verdict A well-ablated empirical win for VLM-generated training data in CD-FSOD, but the central augmentation runs through an undisclosed prompt and a closed API; referee it, then ask for the prompt and error bars. read the letter →

arxiv 2608.01348 v1 pith:SWW4WH6D submitted 2026-08-02 cs.CV

Linhai Zhuo , Junxi Cai , Tianwen Qian , Qingping Zheng , Yang Liu This is my paper

classification cs.CV

keywords cross-domainfew-shotobjectdetectiondataaugmentationvision-languagemodelsprompt-drivenimagegenerationfeatureperturbationdomain-invariantrepresentationGroundingDINO

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

PSP-FSOD claims that a vision-language model can act as a controllable data-augmentation engine for cross-domain few-shot object detection: given a support image and its bounding box, a structured prompt makes the model generate new training images that keep the object's category and structure while shifting its appearance, pose, and background to match the target domain. The paper further claims that injecting Gaussian noise into multi-scale detector features, followed by a mean-variance correction step, stabilizes training and reduces reliance on domain-specific cues. Together these two components raise average mAP on the CD-ViTO benchmark from 33.6 to 36.4 (1-shot), 42.7 to 45.0 (5-shot), and 45.4 to 47.1 (10-shot) over the previous augmentation-based method, Domain-RAG. The result matters because it suggests a cheap, training-free route to synthetic target-domain supervision without acquiring more real labels.

What carries the argument

The central mechanism is prompt-driven domain simulation (PDS), built on a structured prompt $P(y, d) = \{P_{\mathrm{sem}}, P_{\mathrm{fg}}, P_{\mathrm{bg}}, P_{\mathrm{loc}}, P_{\mathrm{ann}}\}$ that separately constrains semantic identity, foreground adaptation, background adaptation, object placement, and annotation output. A grounding-aware generation scheme uses a visible red box as an intermediate cue to preserve spatial correspondence, then removes it from the final image. The complementary mechanism is noise-induced feature perturbation (NFP), which injects Gaussian noise $X^{\mathrm{noisy}}_i = X^{\mathrm{clean}}_i + \sigma_i \epsilon_i$ with $\sigma_1 = 0.15$ for low-level features

What would settle it

Replace Gemini with an open-weight grounded VLM using the same five-part prompt and measure average mAP on the CD-ViTO benchmark; if performance falls back to the Domain-RAG or baseline level, the claim depends on the specific closed model. Alternatively, run the generation protocol on a sample of support images and have human annotators check whether the returned bounding box actually contains the intended object; a high rate of category or bounding-box misalignment would directly undermine the claimed semantic-spatial consistency.

Watch

Extended reading notes

Core claim

The paper's central claim is that the bottleneck in CD-FSOD is not only lack of target data but the type of augmentation used to fake target-domain data: background-only adaptation such as Domain-RAG leaves foreground representations fixed, while naive copy-paste and low-level transforms fail to model complex shifts. PSP-FSOD therefore uses a VLM, Gemini, as a training-free generator. Each support image and bounding box are fed with a five-part structured prompt $P(y, d)$ that (1) constrains semantic category, (2) adapts foreground, (3) adapts background, (4) fixes object placement, and (5) asks for the image plus bounding box, producing $(\hat{I}, \hat{B})$. Generated samples are added to t

Load-bearing premise

The entire gain rests on the closed VLM reliably following the five-part prompt so that generated images preserve the object's category, structure, and location while returning usable bounding boxes; the specific prompt text and generation protocol are not disclosed, so this cannot be independently verified.

Editorial extensions

If this is right

Joint prompt-driven foreground–background generation is the dominant source of gains, adding +6.4 mAP at 1-shot over the fine-tuned GroundingDINO baseline, while feature perturbation adds +1.1 mAP.
The full PSP-FSOD outperforms Domain-RAG by 2.8, 2.3, and 1.7 mAP at 1-, 5-, and 10-shot settings, and outperforms the fine-tuned GroundingDINO baseline by 10.1, 4.6, and 3.2 points.
Gains are largest on domains with strong foreground variation and foreground–background coupling, such as ArTaxOr, DeepFish, UODD, and NEU-DET, and smaller on simpler domains like Clipart1k and DIOR.
The method needs only one or two generated images per support image, fewer than Domain-RAG's five, indicating that alignment and semantic consistency matter more than raw generation volume.
Because the generation stage is training-free and the noise injection is inserted between the backbone and feature enhancer, the authors argue the framework extends to other few-shot learning tasks and detectors built on multi-scale features.

Reading between the lines

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

If the gains come from VLM grounding rather than mere image diversity, the five-part prompt structure is a transferable recipe: the same decomposition into semantic constraints, foreground/background adaptation, location, and annotation output could be packaged as a reusable augmentation protocol for other detection domains.
The success with one generated image per support sample hints that quality and alignment matter more than volume; a testable extension is to vary the number of generated samples and measure whether performance saturates, which would reveal a 'sufficient synthetic coverage' regime.
Because the noise injection uses only mean-variance correction, it can likely be ported to other dense prediction tasks such as segmentation or keypoint detection under few-shot domain shift, though the paper does not test this.
The dependency on a closed VLM means reproducibility hinges on whether any grounded VLM can follow the same prompt; reimplementing the generation with an open-weight model would test whether the mechanism transfers or is model-specific.

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

Couple of things to know. This is a solid empirical paper in the niche CD-FSOD area. The new bit is joint foreground-background prompt-driven synthesis with a grounding-aware placement step, which is a genuine combination relative to Domain-RAG (background-only) and ETS (search-based). The ablations in Table 3 directly compare foreground-only, background-only, and joint, and joint wins on every dataset. The PDS ablation in Table 2 gives +6.4/+3.2/+2.1 mAP over Baseline+, which is the main effect. They also add a noise-induced feature perturbation that has a mean-variance correction; that part is close to standard feature noise with a re-normalization trick, so don't oversell it.

Where it gets soft. The biggest issue is reproducibility. The entire PDS pipeline is Eq. (1): a Gemini call with a five-part prompt P(y,d). The prompt is decomposed into semantic constraints, foreground adaptation, background adaptation, location, and annotation, but the actual text is never shown. Gemini is a closed API. The paper says an anonymized code repo exists, but code can't reproduce the synthesis without the prompt and the API. There's no human rating of semantic consistency and no quantitative check that the returned boxes B-hat are correct, even though the paper's own Fig. 4 shows the failure mode when the prompt doesn't enforce consistency. That makes the central causal claim untestable as reported. This doesn't make the numbers false, but it means the +6.4 from PDS flows through an unverifiable black box.

Second, single-run results, no error bars, acknowledged in checklist item 7. For a method with this much randomness (VLM generation, noise injection, fine-tuning), a few seeds would matter. Third, NFP is not consistently helpful: in the ablation, NFP alone drops DeepFish 1-shot from 42.2 to 36.7, and DIOR/UODD 5-shot by about a point. The paper says 'both modules consistently improve performance,' which is just not true for NFP by itself. Combined they still help, but the sentence overclaims.

Also minor: hyperparameters like epochs and LR milestones are per dataset (Appendix Table 6). That's not disqualifying for few-shot benchmarks, but it is tuning. Self-citations are fine; the cited Domain-RAG is the direct competitor and the paper reproduces its results fairly.

Bottom line: this deserves a serious referee. The core idea is plausible, the ablations are informative, and the gains over Domain-RAG are consistent per-domain. What it needs is a clear disclosure of the prompt, release of the generated dataset or a detailed protocol, error bars, and a more careful claim about NFP. If you're in the CD-FSOD area, read it for the ablation design; don't rely on the exact numbers until the pipeline is reproducible.

Referee Report

4 major / 6 minor

Summary. The paper proposes PSP-FSOD, a method for cross-domain few-shot object detection (CD-FSOD) that combines two components: (a) prompt-driven domain simulation (PDS), which uses the Gemini vision-language model to synthesize domain-diverse support images with joint foreground-background adaptation and grounding-aware bounding-box generation; and (b) noise-induced feature perturbation (NFP), which injects Gaussian noise into multi-scale features of GroundingDINO with a distribution-correction step during fine-tuning. On the CD-ViTO benchmark, the method reports average mAP of 36.4/45.0/47.1 for 1/5/10-shot, surpassing the Domain-RAG baseline by 2.8/2.3/1.7 points. The paper includes ablations of PDS and NFP, prompt-strategy comparisons, noise-strength analysis, and qualitative visualizations.

Significance. If the reported results are reproducible and statistically reliable, the paper would make a useful empirical contribution to CD-FSOD: it is simple, builds on standard GroundingDINO, and addresses both foreground and background variation with a VLM, while adding a lightweight feature-level regularizer. The evaluation uses the established CD-ViTO benchmark and compares against several independent data-augmentation baselines (ETS, Domain-RAG). The ablations are reasonably designed, including prompt-strategy variants and noise-level sensitivity. The paper's main limitations are also acknowledged in the checklist (no error bars) and in the appendix (manual prompt/noise tuning). However, the core PDS component relies on an undisclosed closed-source generation protocol, which currently prevents independent verification of the central claim.

major comments (4)

[§3.1, Eq. (1), Fig. 4, Table 2] The central causal claim—that PDS contributes +6.4/+3.2/+2.1 mAP (Table 2)—flows entirely through an unverifiable black box. The exact structured prompt P(y,d) is never written out; the five components are described only at a high level. The conversion from the intermediate 'red box' to the final annotation B̂ is described informally. Gemini is a closed API whose behavior can change. Fig. 4 itself shows that when the prompt does not enforce foreground-background consistency, the VLM can produce semantically/spatially wrong compositions. The paper must release the exact prompt template, the generation parameters, the post-processing code, and either the generated dataset or a substantial reproducible subset with human-rated semantic/spatial correctness. Without this, the reported attribution of gains to PDS is not independently testable.
[Checklist item 7, Table 1, Table 2] The paper reports single-run results with no error bars, confidence intervals, or significance tests; the checklist explicitly answers 'No' to this item. This is load-bearing because several headline gains over Domain-RAG are modest (2.8, 2.3, and 1.7 mAP on average), and per-dataset differences are often small—e.g., 5-shot DeepFish is 43.9 vs. 43.8 in Table 1. Single runs cannot distinguish real improvements from seed/initialization noise. At minimum, multi-seed mean±std should be reported for the main results and for the core ablations (Table 2), or the paper should justify why single runs are sufficient under this protocol.
[Table 2 (NFP rows), §4.3] The text claims that 'both modules consistently improve performance,' but the NFP-only ablation does not consistently improve: on DeepFish 1-shot, NFP reduces mAP from 42.2 to 36.7 (−5.5 points); on DIOR 5-shot it is 26.9 vs. 27.9; on UODD 5-shot it is 26.7 vs. 27.7. Even in the combined model, 10-shot NEU-DET drops from 28.3 (PDS alone) to 27.8 (Ours). The NFP contribution is therefore positive mainly when combined with PDS. The paper should report per-domain NFP effects honestly and avoid the 'consistently improves' wording, or provide a mechanism explaining these negative cases.
[§4.1, Table 5, Table 6, Appendix C] The selection of NFP hyperparameters is under-specified. Table 5 tunes noise strengths only on UODD, yet the method section fixes σ1=0.15 and σ2=0.10. The implementation details say 'Dataset-specific settings, including noise strengths... are provided in the Appendix,' but Table 6 lists only epochs, batch sizes, and LR decay milestones; no dataset-specific noise-strength table is given. This is an internal inconsistency and a reproducibility gap. Please report the noise strengths actually used per dataset, and add a sensitivity analysis beyond UODD if the claim is that the same fixed values transfer.

minor comments (6)

[Eq. (1), §3.1] Define all notation: G, I, B, d are used without a formal description of their spaces. The phrase 'visible red box' appears only in text; clarify how B̂ is computed from the red box and whether the generated image is post-processed before training.
[Table 1] The Distill-CDFSOD row contains 'nan' for DeepFish and NEU-DET without explanation. Also, 'best results highlighted in pink' is not accessible in monochrome printing; use bold or a table note.
[Tables 2 and 3] The 'Baseline+' of Table 2 and the 'GroundingDINO' row of Table 3 differ (e.g., ArTaxOr 26.8 vs. 26.3), but the difference is not explained. Define both baselines explicitly so readers can interpret the ablations.
[§3.2 vs. Table 5] The terminology is inconsistent: the method section says 'distribution correction,' while Table 5's caption says 'distribution rectification.' Use one term throughout.
[Checklist item 5, §4.1] The checklist states that an anonymized code repository is provided, but no URL or repository identifier appears in the manuscript. Either include the URL or state that code will be released upon publication.
[Throughout] Minor typos and phrasing: 'aprompt-driven' should be 'aprompt-driven' (Sec. 1); 'generated figures bring larger gains' likely means 'generated samples/images' (Sec. 4.3); 'L20X GPUs' should be specified with memory and total compute if possible.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: gains are measured against external benchmarks and held-out test labels; PDS is a training-time augmentation pipeline, not a fit to the target result.

full rationale

The paper's central claims are empirical: it reports mAP on the CD-ViTO benchmark under 1/5/10-shot protocols and compares against external baselines (GroundingDINO, ETS, Domain-RAG). The PDS module uses support images, their bounding boxes, and a category label to synthesize additional training samples (Eq. 1), which are then used to fine-tune GroundingDINO. Evaluation is performed on a disjoint query set with human-annotated labels, so the reported improvements are not forced by the generation process by construction. The NFP module injects Gaussian noise and then rescales perturbed features to match the clean feature statistics; that matching is definitional, but it is a design property of the regularizer, not a derived result that is then 'predicted'. The paper's limitations—manually designed prompts, manually selected noise strengths, no error bars, and an undisclosed prompt template—are reproducibility and robustness concerns, not evidence of circularity. There is co-author overlap with Domain-RAG (Tianwen Qian), but Domain-RAG is treated as a competitor baseline rather than as load-bearing justification for the method, so this does not raise the circularity score. No step in the derivation chain reduces to its own input or to a self-citation chain.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework introduces no new physical entities or latent variables; the contributions are an empirical augmentation pipeline and a feature regularizer. The listed free parameters are hand-set or per-dataset choices, and the axioms are assumptions about VLM fidelity and the transferability of noise regularization.

free parameters (5)

Noise strengths σ1, σ2 = σ1=0.15, σ2=0.10
Chosen by hand through ablations (Table 5), per feature-level group, not derived.
Number of generated images per support image = 2 for ArTaxOr, 1 for others
Selected per dataset, Appendix A.2, to cover diversity.
Training schedule (epochs, batch size, LR milestones) = e.g., epochs 30-80, batch 3-4, milestones per dataset
Dataset-specific hyperparameters in Table 6, no validation protocol described.
Injection probability = 1
Noise injected every forward pass, fixed by design.
Prompt template P(y,d) = manual
Prompt template manually designed, acknowledged in Appendix C as a limitation; text not shown in paper.

assumptions (4)

domain assumption Gemini-generated images preserve semantic category and annotation alignment
Section 3.1 relies on VLM grounding to produce usable (Î, B̂); the paper's own Fig. 4 shows mismatch failures when prompts are inconsistent.
domain assumption Gaussian noise with mean-variance correction yields domain-invariant features
Section 3.2 heuristic; ablations show mixed results (e.g., DeepFish 1-shot drops 42.2 to 36.7, Table 2).
domain assumption Target-domain description d can be obtained by an LLM from dataset papers
Section 3.1, d is not specified or evaluated.
domain assumption Standard supervised fine-tuning objectives of GroundingDINO transfer to augmented samples
Section 3.3, augmented samples are treated as real labels without uncertainty weighting.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Prompt-Driven Simulation with Feature Perturbation for Cross-Domain Few-Shot Object Detection." pith.science (2026). https://pith.science/paper/SWW4WH6D

@misc{pith2026260801348,
  author       = {Pith},
  title        = {Pith review of: Prompt-Driven Simulation with Feature Perturbation for Cross-Domain Few-Shot Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SWW4WH6D}},
  note         = {Machine review of arXiv:2608.01348}
}

read the original abstract

Data augmentation, which simulates diverse visual variations to expand the source distribution and induce synthetic domain shifts, is a simple yet effective strategy for mitigating severe domain shifts and limited labeled target data in cross-domain few-shot object detection (CD-FSOD). Existing approaches rely on conventional data augmentation, such as Color-Jitter, Mosaic, and background-centric adaptation (e.g., Domain-RAG), which are limited in modeling complex domain shifts and often lead to suboptimal performance. In this paper, we propose PSP-FSOD, a principled framework that integrates prompt-driven domain simulation with feature perturbation regularization to improve generalization in CD-FSOD. To enable controllable domain synthesis, we design a prompt-driven strategy that leverages the visual grounding capability of large VLMs to jointly model foreground and background variations, generating semantically consistent yet domain-diverse training samples. Moreover, we adopt a grounding-aware generation scheme that guides object placement and alleviates semantic-spatial misalignment, thereby improving foreground adaptation. To ensure training stability and robustness, we further introduce a noise-induced feature perturbation mechanism that injects Gaussian noise into multi-scale intermediate features with distribution correction, encouraging consistent predictions under perturbations and reducing reliance on domain-specific cues. Extensive experiments demonstrate that PSP-FSOD produces high-quality domain-diverse supervision and learns domain-invariant representations, consistently improving performance across CD-FSOD benchmarks.

Figures

Figures reproduced from arXiv: 2608.01348 by the authors.

**Figure 2.** Overview of the proposed framework. (a) Prompt-Driven Domain Simulation generates [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Visualization of different components on target-domain images under the 1-shot setting. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Visualization comparison of different prompt-driven domain simulation strategies. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

**Figure 5.** Figure 5: Visualization comparison of different methods on target-domain images. “Baseline+” is [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

**Figure 6.** Figure 6: Additional detection results of PSP-FSOD on diverse target-domain images. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

**Figure 7.** Figure 7: Additional generation results of PSP-FSOD. Bounding boxes are overlaid only for [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

**Figure 8.** Figure 8: Failure cases of PSP-FSOD on challenging target-domain images with severe domain [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

The First EgoCross Challenge at EgoVis 2026: Cross-Domain Egocentric Video Question Answering
cs.CV 2026-08 accept novelty 5.0 of 10

At CVPR 2026, the EgoCross Challenge evaluated cross-domain egocentric video QA across four specialist domains, with the best system reaching 66.98% accuracy.

Reference graph

Works this paper leans on

16 extracted references · 6 canonical work pages · cited by 1 Pith paper

[2]

Qwen3 technical report.arXiv preprint arXiv:2505.09388,

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

arXiv
[3]

Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalk- wyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

arXiv
[6]

mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412,

Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412,

arXiv
[8]

Lei Qi, Hongpeng Yang, Yinghuan Shi, and Xin Geng

doi: 10.1109/ TCSVT.2022.3152615. Lei Qi, Hongpeng Yang, Yinghuan Shi, and Xin Geng. Normaug: Normalization-guided augmentation for domain generalization.IEEE Transactions on Image Processing,

work page arXiv 2022
[9]

Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang

doi: 10.1109/TIP.2024.3364516. Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Wave-san: Wavelet based style augmentation network for cross-domain few-shot learning.arXiv preprint arXiv:2203.07656,

work page arXiv 2024
[11]

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo

doi: 10.1109/CVPR52733.2024.01682. Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012–10022,

arXiv 2024
[14]

DeFRCN: Decoupled Faster R-CNN for Few-Shot Object Detection

Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection.arXiv preprint arXiv:2108.09017,

work page Pith review arXiv
[15]

Detect everything with few examples.arXiv preprint arXiv:2309.12969,

Xinyu Zhang, Yuhan Liu, Yuting Wang, and Abdeslam Boularias. Detect everything with few examples.arXiv preprint arXiv:2309.12969,

arXiv

Show all 16 references

[16]

A.1 More Implementation Details Additional implementation details are summarized in Table 6, including dataset-specific epochs, batch sizes, and learning rate decay milestones

A More Implementation Details and Analysis This section presents additional implementation details of the proposed CD-FSOD method and comparison of the number of the generated samples. A.1 More Implementation Details Additional implementation details are summarized in Table 6,...

2019
[2019]

Frustratingly simple few-shot object detection.arXiv preprint arXiv:2003.06957,

Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection.arXiv preprint arXiv:2003.06957,

2003 arXiv
[2020]

Raphael Gontijo Lopes, Dong Yin, Ben Poole, Justin Gilmer, and Ekin D

doi: 10.1109/W ACV45572.2020.9093467. Raphael Gontijo Lopes, Dong Yin, Ben Poole, Justin Gilmer, and Ekin D. Cubuk. Improving robustness without sacrificing accuracy with patch gaussian augmentation.arXiv preprint arXiv:1906.02611,

2020
[2021]

Arthropod taxonomy orders object detection dataset

Geir Drange. Arthropod taxonomy orders object detection dataset. https://doi.org/10.34740/kaggle/ dsv/1240192,
[2022]

Cd-fsod: A benchmark for cross-domain few-shot object detection

Wuti Xiong. Cd-fsod: A benchmark for cross-domain few-shot object detection. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE,

2023
[2023]

Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

arXiv
[2024]

Andreas Bär, Neil Houlsby, Mostafa Dehghani, and Manoj Kumar

doi: 10.1145/3696409.3700174. Andreas Bär, Neil Houlsby, Mostafa Dehghani, and Manoj Kumar. Frozen feature augmentation for few- shot image classification. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16067–16077,
[2025]

Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952,

Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952,

arXiv

Pith tools

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

[1] [2]

Qwen3 technical report.arXiv preprint arXiv:2505.09388,

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

arXiv

[2] [3]

Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalk- wyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

arXiv

[3] [6]

mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412,

Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412,

arXiv

[4] [8]

Lei Qi, Hongpeng Yang, Yinghuan Shi, and Xin Geng

doi: 10.1109/ TCSVT.2022.3152615. Lei Qi, Hongpeng Yang, Yinghuan Shi, and Xin Geng. Normaug: Normalization-guided augmentation for domain generalization.IEEE Transactions on Image Processing,

work page arXiv 2022

[5] [9]

Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang

doi: 10.1109/TIP.2024.3364516. Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Wave-san: Wavelet based style augmentation network for cross-domain few-shot learning.arXiv preprint arXiv:2203.07656,

work page arXiv 2024

[6] [11]

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo

doi: 10.1109/CVPR52733.2024.01682. Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012–10022,

arXiv 2024

[7] [14]

DeFRCN: Decoupled Faster R-CNN for Few-Shot Object Detection

Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection.arXiv preprint arXiv:2108.09017,

work page Pith review arXiv

[8] [15]

Detect everything with few examples.arXiv preprint arXiv:2309.12969,

Xinyu Zhang, Yuhan Liu, Yuting Wang, and Abdeslam Boularias. Detect everything with few examples.arXiv preprint arXiv:2309.12969,

arXiv