Thank you very much for the positive and constructive feedback! We really appreciate that you consider our work “easy to follow”, “training-free”, “can synthesize high-quality samples”, and delivering “state-of-the-art results and interesting ablations”. Below, we address the concerns raised.

W1&Q5: Potential General Usage Instead of being Tailored for CD-FSOD.

Thanks. We appreciate this constructive suggestion. However, we would also like to highlight that:

1. We focus on CD-FSOD mainly due to its unique and challenging characteristics: significant domain shift and limited training examples. These two challenges make almost all the prior image generation can’t work well. In contrast, our method perfectly addresses their limitations with our training-free, retrieval-based augmentation strategies, showing a clear performance boost over other methods. That said, if this is not the case with CD-FSOD, we most likely could propose new solutions with better solutions, for example, finetuning when the examples are not few-shot.

2. Besides, in addition to the CD-FSOD it alone, in our paper, we also considered and conducted experiments on other FSOD tasks: in Tab.2, we studied the standard Remote-Sensing FSOD (which falls within general FSOD instead of CD-FSOD), and the adapted Cross-Domain Remote-Sensing FSOD; in Tab.3, we studied the CamouflageFSOD. These broader experiments demonstrate our effectiveness beyond CD-FSOD.

For more broader scope, e.g., general object detection, we agree our method has potential usage but also we think the current work might not the optimal form, so we will take the suggestion and explore it as our future work.

W2: Experiment Details and Fairness Due to Data Usage and Models.

Thanks. For the experiment details, in addition to Setups and Implementation Details provided in Sec.4 (L204-L220), we also provide more details in Sec.A.1(Appendix), including more details on the proposed method (Sec.A.1.1), more details on training (Sec.A.1.2), and more ablation studies (Sec.A.1.3).

For the Fairness issue, on the data side, we clarify that: though we use COCO as a database to augment new images, COCO itself is defined as the source in the default CD-FSOD setting, and we don’t break this formulation. It fundamentally should be regarded as different ways of using the source data. Additionally, 1) the comparison between ours (33.6, Tab.1) and other competitors, e.g., ETS (27.8, Tab.1), Copy-Paste (27.9, Tab.4), Foreground Aug. (28.1, Tab.4), and Background Aug. (29.5, Tab.4) is fair: we share the exact same baseline and training configuration, and all of them are data augmentation-based approaches. For example, ETS proposes a mixture of 6 various augmentation strategies during training; 2) the ablation experiments on Fig.3(a) and Tab.5 (Appendix) all use COCO as DB while without applying the complete Domain-RAG method. We believe these results confirm that the improvements stem from our pipeline design rather than merely the introduction of synthetic data.

On the model side, since we aim to build a training-free augmentation, the introduction of extra generation models, e.g., FLUX, is unavoidable. However, we highlight: 1) As stated before, the baseline (GroundingDINO) is the same as others; 2) All the extra introduced models, e.g., FLUX, are only used for the generation; during the inference time, only the single base GroundingDINO is adopted, which is also the same as others.

W3&Q1&Q3: Comparison of Model Efficiency, such as Parameters and Inference Speed.

Thanks, our method builds on GroundingDINO without modifying any model architecture or training/inference process on the target support sets. Therefore, both parameter count (233.00 M) and inference speed (~0.22 seconds per image) are the same as those of competitors that are built on the same GroundingDINO, e.g., GroundingDINO itself, EST, and our Domain-RAG.

In addition, upon request, we further provide the computation time needed for our image generation below. We emphasize that: 1) the generation indeed takes time, while it is a long-standing issue in the whole generation community. 2) The generation period introduces no additional cost during inference. 3) The current bottleneck of CD-FSOD lies in the model performance, due to its huge challenges, while our method provides practical and effective solutions for enhancing existing detectors for CD-FSOD.

These together make our method still scalable, especially under low-shot cases, where only a few query images are needed to be augmented.

Q2: Grounding-DINO with/without Fine-tuning.

Thanks, we provide a comparison of GroundingDINO under different training strategies: zero-shot, 1-shot fine-tuning, and with Domain-RAG.

We notice that:

• Zero-shot GroundingDINO shows limited generalization, especially on challenging datasets (e.g., ArTaxOr, NEU-DET).
• Fine-tuning with 1-shot improves results moderately.
• Domain-RAG with fine-tuned GroundingDINO achieves consistent gains across all datasets, with a +30.9 mAP increase on ArTaxOr and +4.3 mAP on UODD compared to 1-shot GroundingDINO.

This demonstrates that GroundingDINO is a strong baseline and Domain-RAG helps improve its generalization in few-data, cross-domain scenarios further.

Q4: A Stronger Backbone Used.

As mentioned in W2, we conducted controlled experiments using the same backbone and training setup as competitors (e.g., ETS) to ensure a fair comparison.

Domain-RAG consistently outperforms these baselines, showing that the performance gain comes from our augmentation method rather than differences in model capacity. To further demonstrate the generality of Domain-RAG, we applied it to DEViT-FT and CD-ViTO. As shown below, both “DEViT-FT + Domain-RAG” and “CD-ViTO + Domain-RAG” yield consistent improvements across all datasets.

These results confirm that Domain-RAG improves detection performance across different model architectures, supporting its applicability as a general-purpose augmentation strategy. We will include these in the revised version.

We hope that our response adequately addresses your concerns. We are happy to provide more information during the discussion phase.

Thank you very much for the positive and constructive feedback! We really appreciate that you consider our work “creative”, “effective”, “significantly improving background realism and coherence”, and with “clear state-of-the-art results” . Below we address the concerns raised.

W1: Primarily on Empirical Findings without Rigorous Theoretical Analysis.

Thanks, we agree that our paper focuses on empirical contributions rather than theoretical analysis. While this is the case, our work targets a highly challenging and practical problem, cross-domain few-shot object detection, motivated by the limitations of existing approaches, which either require retraining, rely on domain-specific designs, or lack scalability. Facing such a challenging task, we contributed a novel, effective, training-free, model-agnostic framework that addresses image generation across domains through the principle of “fix the foreground, adapt the background.” This idea is not only conceptually new but also practically significant.
Furthermore, we would like to highlight though without mathematical analysis, we extensively validate our method through:

• Effective and widely validated experiments across diverse benchmarks (CD-FSOD, RS-FSOD, Camouflaged FSOD), and across 8 distinct target domains.
• Comprehensive and detailed ablation studies and analysis demonstrating the contribution of each component.
• Completely training-free, enabling easy integration into existing detection pipelines without additional overhead.

We believe this combination of innovative design, strong empirical results, and practical feasibility also provides valuable insights and a robust solution for the community, complementing theory-driven research.

W2: Dependence on Quantity of Retrieval DB and Generalization.

We acknowledge the concern and would like to clarify that the domains we tested already include drastically different scenarios, such as fine-grained insects (ArTaxOr), cartoon-style drawings (Clipart1k), aerial remote-sensing imagery (DIOR, NWPUVHR-10), underwater cases (UODD, DeepFISH), and industrial textures (e.g., NEU-DET), which are far from the natural image distribution of COCO. Despite this large gap, our method consistently outperforms strong baselines, other competitors, demonstrating robustness even when the retrieval database and target domain differ significantly.

Furthermore, upon question, we conducted additional experiments under three conditions: 1) COCO as DB but with reduced category numbers, 2) NEU-DET (non-general domain) as DB, and 3) also miniImageNet as DB (originally reported in Tab.6, Appendix)

From the below results, we show that: 1) more broader category coverage improves performance, but even with only 1 COCO category, we still clearly enhance the results. 2) Even when the source domain (e.g., NEU-DET) is far from COCO, Domain-RAG still delivers strong gains through its retrieval-conditioned generation and style adaptation modules, 3) As discussed in L927-L935, mini-ImageNet also works as DB, which is also easy to obtain.

From these results, we could conclude that our Domain-RAG enhances the baseline performance regardless of the database, solidly validate our effectiveness. And, the selection of more general-purpose domains e.g., COCO and ImageNet are widely available, thus require no special design, making our approach practical and scalable.

Q1: Additional Controls or Safeguards to Mitigating the Foreground Leakage?

Thanks, we appreciate the discussion. We also considered this question ourselves during the method development and Reviewer-wq5x (Q3) also asked about this. Here are our responses:

Both we authors and Reviewer-wq5x consider adding post-filtering as a safeguard; however, it is non-trivial to achieve a good and especially efficient filtering strategy. Specifically,

1. Directly applying ZSL inference of existing detectors for foreground detection on CDFSOD is weak, making it difficult for a pseudo-detector to accurately identify leakage. While if we want to apply the trained detector, e.g., trained Domain-RAG (based on GroundingDINO), it will largely burden the pipeline to: data augmentation → train model → filter → train model again.

2. We also consider applying VLMs (e.g., GPT-4o) to help with filtering. However, they also struggle to meet our expectations. In the meantime, it will also increase the time cost and complexity.

Besides, we analyzed leakage cases and found that among all generated images in the CDFSL, for example, only very few examples exhibited a foreground leakage issue. For example, 1.3% (4/290) under 1shot.

Considering these trade-offs, and given the extremely low frequency of leakage (1.3%), we decided not to introduce post-filtering as it would not significantly impact overall performance while violating the method’s training-free and lightweight design.

Q2: Adapt to Zero-Shot Scenarios?

Thanks. Our method assumes access to target-domain foreground samples, which are provided in any few-shot tasks. Fully zero-shot scenarios, where no target images are available, fall outside this scope. We will better clarify in the revised version.

Q3: Selection of Hyperparameter (λ₁, λ₂).

The 1-shot results of different blending coefficients (λ₁, λ₂) are provided below. Experiments show that λ₁ = 1 and λ₂ = 0.8 achieve the best balance. When λ₂ = 1, we find that λ₁ = 1 performs best. During testing, we slightly reduced λ₂ to 0.8 because setting it too low (e.g., λ₂ = 0.5) distorts semantics and degrades generation quality. This design ensures that domain-specific cues are preserved while still incorporating context from the retrieval database.

Q4: Caption of Figure 3 contains a typo. (a) → a).

Thanks for your careful review. We will fix it!

Q5: Effect of Augmented Sample Count per Category.

Thanks, we conducted the ablation of different numbers of augmented samples per category (n) in Tab.8, Appendix (where “m” was incorrectly written instead of “n”; this will be corrected). n is selected as 5 according to the ablation results. Note that once we select the best n, we keep it the same for all 3 tasks and 8 targets, without manually tuning it on each target.

We hope that our response adequately addresses your concerns. We are happy to provide more information during the discussion phase.

Thank you very much for the positive and constructive feedback! We really appreciate that you consider our work “easily adaptable”, “motivated by shortcomings of existing methods”, “extensively validated”, and with “ realistic image generation” . Below we address the concerns raised.

W1: Ablation on Synthetic Sample Quantity n.

Thanks for pointing this out. We actually provided an ablation study on the number of generated samples in Tab.8, Appendix (where “m” was incorrectly written instead of “n”; this will be corrected). Results show that performance improves significantly as n increases from 1 to 3, achieves best beyond n=5, and drops when n is set as 10. As in L941-950, we analysis that too few retrieved images limit visual variation, while too many increase the chance of unrealistic compositions.

W2: Fairness and Generalizability in Experimental Comparisons.

Thanks, we first clarify the reason for building our model on GroundingDINO, discuss more on the fairness, and then provide more results regarding the generalizability.

• Reason for building upon GroundingDINO. The primary reason we report results on GroundingDINO is due to its superior performance compared to other detectors (GroundingDINO, 26.3, 1shot, Tab.1) vs. (CD-ViTO, 13.9, 1shot, Tab.1) of average mAP, which we believe will advance this area rapidly.
• Fairness of Comparison Regarding the Usage of Synthetic Data. First of all, though we use COCO as database to augment new images, COCO itself is defined as the source in the default CD-FSOD setting, we don’t break this formulation. 2) The comparison between ours (33.6, Tab.1) and other competitors e.g., ETS (27.8, Tab.1), Copy-Paste (27.9, Tab.4), Foreground Aug. (28.1, Tab.4), and Background Aug. (29.5, Tab.4) are fair: we share the exact same baseline and training configuration, and all of them are data augmentation based approaches. For example, ETS proposes a mixture of 6 various augmentation strategies during training; The comparison between ours and CD-ViTO and Detic-FT is unfair, but we clearly notice that we use different baselines in Tab.1, and we are not using these performance gap to show our effectiveness. 3) Additionally, the ablation experiments on Fig.3(a) and Tab.5 (Appendix) all use COCO as DB while without applying the complete Domain-RAG method. We believe these results confirms that the improvements stem from our pipeline design rather than merely the introduction of synthetic data.
• Generalizability of Domain-RAG to Other Detectors. In addition to building on top of GroundingDINO, we also demonstrated our results on SAE-FSDet, the prior SOTA method on remote sensing FSOD, as in Tab.2. Upon question, we further tested Domain-RAG on other detectors on DeViT-FT and CD-ViTO. Below results demonstrate that Domain-RAG is detector-agnostic and can be seamlessly integrated into different architectures without requiring retraining or architectural modifications.

W3 & Q1: Reliance on the Diversity of the Background Database (COCO).

Thanks, we appreciate this discussion. We would like to clarify that the tested domains are far from being COCO-like. They already involve extremely challenging scenarios with significant category shifts (e.g., novel classes unseen in COCO), severe domain shifts (e.g., clipart-style images, aerial imagery), and also non-natural settings (e.g., industrial textures in NEU-DET).

To further validate the impact of database diversity and bias, we conducted additional experiments under three conditions: 1) COCO as DB but with reduced category numbers, 2) NEU-DET (non-general domain) as DB, and 3) also miniImageNet as DB (originally reported in Tab.6, Appendix)

From the below results, we show that: 1) more broader category coverage improves performance, but even with only 1 COCO category, we still clearly enhance the results. 2) Even when the source domain (e.g., NEU-DET) is far from COCO, Domain-RAG still delivers strong gains through its retrieval-conditioned generation and style adaptation modules, 3) As discussed in L927-L935, mini-ImageNet also works as DB, which is also easy to obtain.

From these results, we could conclude that our Domain-RAG improves the base groundingDINO regardless of the database, solidly validate our effectiveness.

W3: Analysis on Performance Drop at Higher Shot Settings.

Thanks for pointing this out. We conducted further analyze to understand the slight mAP drop observed at higher shot counts (e.g., 20-shot in RS-FSOD, Tab.2). Specifically, we varied the number of generated samples (n) under the 20-shot setting in RS-FSOD. Results on VHR-10 are as below. Results validate that reducing n alleviates performance degradation at high shot counts (e.g., 85.48% with n=1 vs. 84.15% with n=3), which better validate our guess in L256 to L257. We highlight that: 1) This effect is limited to high-shot scenarios and can be addressed with a simple adjustment. 2) Domain-RAG only drops in this single case without manually tuning hyperparameter according to the targets, and achieves consistent improvements in 1-, 5-, and 10-shot settings across 3 tasks and 8 distinct target domains.

Q2: Hyperparameters: Retrieval (m, n), blending coefficients (λ₁, λ₂).

Thanks, 1) the results of different n are answered in W1, which was provided in Tab.8, Appendix; 2) m is set as 100 (L215), the model is robust to m since as long as it covers n (e.g.,5) images it doesn’t affect the results in principle; 3) the 1shot results of different blending coefficients (λ₁, λ₂) are provided below. Experiments show that λ₁ = 1 and λ₂ = 0.8 achieve the best balance. When λ₂ = 1, we find λ₁ = 1 performs best. During testing, we slightly reduced λ₂ to 0.8 because setting it too low (e.g., λ₂ = 0.5) distorts semantics and degrades generation quality. This design ensures that domain-specific cues are preserved while still incorporating context from the retrieval database.

We will add these results to the Appendix.

Q3: Could Post-Filtering be Introduced to Address Foreground Leakage/Failure Cases?

Thanks for the suggestion. We also considered this ourselves, however, we ultimately decided not to adopt it for the following reasons:

1) we analyzed leakage cases and found that among all generated images in the CDFSL, only very few examples exhibited foreground leakage issue. For example, 1.3% (4/290) under 1shot.

2) While post-filtering strategies could be applied, they present significant challenges:

• Directly applying ZSL inference of existing detectors for foreground detection on CDFSOD is weak, making it difficult for a pseudo-detector to accurately identify leakage. While if we want to apply the trained detector, e.g., trained Domain-RAG (based on GroundingDINO), it will largely burden the pipeline to: data augmentation → train model → filter → train model again.
• We also consider applying VLMs (e.g., GPT-4o) to help with filtering. However, they also struggle to meet our expectations. In the meantime, it will also increase the time cost and complexity.

Considering these trade-offs, and given the extremely low frequency of leakage (1.3%), we decided not to introduce post-filtering as it would not significantly impact overall performance while violating the method’s training-free and lightweight design.

We hope that our response adequately addresses your concerns. We are happy to provide more information during the discussion phase.

Thanks to the authors' reply. Most of my concerns have been addressed. I decide to raise the rating to borderline accept.

Thank you very much for the positive and constructive feedback! We really appreciate that you consider our work “well-designed”, “seamless and natural generation”, “easy-to-adapt”, “performed well”, and that our writing “well organized and easy to follow”. Below, we address the concerns raised.

W1-1: More Discussion on the Quality and Construction of RAG DB.

Thanks. We first clarify the reasons for choosing COCO as DB and then discuss more on how the quantity of DB affects the results.

• Reasons for taking COCO as DB. In the defined (closed-source) CD-FSOD setting as proposed in CD-ViTO[10], COCO serves as the only single source data for training, while the others (ArTaxOr, Clipart1k, DIOR, DeepFish, NEU-DET, UODD) work as unseen targets. Using COCO as the RAG DB: 1) does not introduce extra data beyond the default setting, protecting the fairness of comparison. 2) As in L49, we hope to generate diverse backgrounds for the query image, which COCO fits well with diverse categories and general-domain backgrounds, which can better cover novel domains.
• How Quantity of DB Affects Results. Upon question, we further conduct more experiments with different DB options, including COCO as DB but with reduced category numbers, NEU-DET (non-general domain) as DB, and also miniImageNet as DB (originally reported in Tab.6, Appendix)

From the results below, we show that: 1) broader category coverage improves performance. 2) More general-domain DB (COCO) works better than specific-domain (NEU-DET). 3) As discussed in L927-L935, mini-ImageNet also works as DB, will be inferior to COCO preliminary due to its larger foreground, i.e., smaller background. We highlight that our Domain-RAG improves the base GroundingDINO in all the cases, and solidly validates our effectiveness.

W1-2: Fairness of Using RAG DB.

Thanks, we will clarify that:

1) As we mentioned in the reason for taking COCO as DB, it is defined as the source in the default CD-FSOD setting; thus, using COCO should be regarded as a different way of leveraging the same source data. 2) Several compared methods include ETS (27.8, Tab.1), Copy-Paste (27.9, Tab.4), Foreground Aug. (28.1, Tab.4), and Background Aug. (29.5, Tab.4), don’t use COCO in a RAG-inspired way, but share the exact same baseline and training configuration with us, and also fall into the data augmentation-based approaches. The performance gap between them and our Domain-RAG (33.6, Tab.1) on 1-shot instead validates the superiority of our RAG-based method, which has not been proposed in this task before. 3) Additionally, the ablation experiments on Fig.3(a) and Tab.5 (Appendix) all use COCO as DB without applying the complete Domain-RAG method. We believe these results confirm that the improvements stem from our pipeline design rather than merely the use of COCO as an external database.

We will include the new results and discussions in the revised paper.

W2-1: Ablation on Inpainting with/without RAG DB as Condition.

Thanks. In Fig. 3(a) and Tab.5 (Appendix), we reported results for w/o Background Retrieval (random COCO background), which shows only limited improvement (31.8 vs Ours: 33.6, 1-shot). To further address your concern, we added experiments for w/o RAG DB (no DB conditioning, where the original image background is used during Domain-Guided Background Generation). We highlight: 1) the comparison results between “w/o RAG DB” vs. “w/o Background Retrieval” or our Domain-RAG indicate that the DB conditioning contributes significantly to final performance, bringing background diversity into the extremely limited target images. 2) The comparison results between “w/o Background Retrieval” vs. our Domain-RAG demonstrate that our domain-aware background retrieval stage provides backgrounds that are better aligned with the target domain (as in L308-309, L922).

W2-2: Discussion on Retrieval Performance and Retrieved Image.

We appreciate this point and thus evaluate the relevance between retrieved samples and its query through human judgment under the 1-shot setting. For most datasets, over 80% of retrieved samples were deemed relevant, except for NEU-DET, where the relevance was lower due to the lack of similar backgrounds in COCO. We also attempted to use GPT-4o for relevance evaluation, but it struggled with context-specific reasoning (e.g., for insect datasets, GPT-4o considered leaves and grass as different contexts, despite both being suitable as natural backgrounds).

Furthermore, we 1) provided the examples of the retrieved background in Fig.5 (Appendix, 2nd column), it shows we managed to retrieval domain-consistent, semantic aligned backgrounds, and 2) again the results summarized in W2-1 between “w/o Background Retrieval (random COCO background) (31.8)” and ours (33.6) well validate that our retrieval images perform well for subsequent image generation.

W2-3: Less of Ablation on the Generated Sample Number n.

Thanks, we actually provided an ablation study on the number of generated samples in Tab.8, Appendix (where “m” was incorrectly written instead of “n”; this will be corrected). Results show that performance improves significantly as n increases from 1 to 3, achieves its best beyond n=5, and drops when n is set to 10. As in L941-950, we analyze that too few retrieved images limit visual variation, while too many increase the chance of unrealistic compositions.

W3: Potential Optimization of the Pipeline.

Thanks for recognizing our pipeline as interesting and proposing this suggestion. While our method builds upon prior diffusion-based inpainting and detection backbones, its key novelty lies in the training-free principle of “fix the foreground, adapt the background” and the carefully integrated pipeline design for addressing the cross-domain image generation. We also considered optimization, for example, applying efficient finetuning, e.g, LoRA, to better fit into targets. However, given extremely few-label examples, e.g,1 shot and also with the domain shift issue, finetuning the model, no matter the number of learnable parameters, doesn’t help improve performance and will burden the current procedure, limiting the quick deployment of our method. In contact, our model-agnostic, training-free approach provides a practical and broadly applicable solution.

Q1: Foreground Object Selection.

Thanks, since we have the object bounding boxes for those support images, i.e., query images to be augmented, we simply fix the objects labeled by those bounding boxes.

Q2: Performance When the Test Sample is Different from the RAG DB.

Thanks, we appreciate this question, while we would like to highlight that: 1) our used CD-FSOD benchmark already include the challenging cases where the target domain is far away from the RAG-DB (COCO), covering domains such as fine-grained animals, cartoon-style, remote sensing, underwater, and industry. For instance, DIOR is a remote sensing image, which is different from the COCO in the visual domain, resolution, camera perspective, and objects. UODD is a more difficult case, containing industry images in gray color. Examples can be found in Fig.3. 2) Even taking NEU-DET, a domain with an extremely large gap with all the others, as the RAG DB, the table below shows we still increase performance over baseline GroundingDINO.

We hope that our response adequately addresses your concerns. We are happy to provide more information during the discussion phase.