acceptodds
Under review as a conference paper at ICLR 2027

What Should We Generate for UAV Detection?

Abstract

Generative augmentation expands the environmental coverage for small unmanned aerial vehicle (UAV) detection, but semantically distinct conditions can lead to similar changes in the detector's features. We combine a two-stage decoupled generation method with detector-guided condition selection. To provide a shared target reference for comparing conditions, we first edit the background and then blend the source UAV back into the image using a soft mask. A frozen detector, trained on clean images, measures paired context responses to estimate condition similarities, and a backward greedy algorithm balances response coverage against pairwise overlap, selecting conditions for a user-specified weight without prescribing the subset size. The retained generated images are combined with clean images for detector training. With five candidate conditions, all selected two- to four-condition configurations achieve higher external mean AP than Clean-only and Full generation with both RT-DETRv2-S and YOLO11m. For RT-DETRv2-S, a selected three-condition subset improves external mean AP by 2.95 points over Clean-only, 0.80 over Full generation, and 2.11 over Full generation with a matched clean/generated sampling ratio. For YOLO11m, the selected subsets improve external mean AP over Full generation by 1.05-1.14 points.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
UAV detection, Generative data augmentation, Detector-guided condition selection, Adverse conditions

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KFSVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a detector-guided generative augmentation framework for UAV detection under adverse conditions. The method first generates different environmental conditions while preserving the original UAV through a two-stage generation process. It then measures feature changes using a frozen detector to estimate similarities between conditions. Based on these similarities, a greedy selection algorithm identifies a subset of conditions that balances coverage and redundancy. The experiments demonstrate that selecting a subset of training set can improve the detection performance compared to using all datasets.

Critical strengths
  1. This paper addresses a practical problem. Generating more diverse conditions does not always improve detection performance. The proposed method seems simple and easy to understand and implement.

  2. The authors test their method on two different detectors, RT-DETRv2-S and YOLO11m. This helps show whether the improvements generalize across different detection models.

Critical weaknesses
  1. Insufficient validation of the proposed condition selection method. The paper shows that selecting a subset of conditions can improve detection performance compared to using all conditions. However, the relationship between condition similarity and detection performance is not sufficiently analyzed. It is unclear how performance changes as more conditions are added, and whether any improvements come from having more training data or greater diversity in augmentation conditions. The proposed method is also not directly compared with simpler alternatives such as random condition selection. Since the experiments only consider five predefined conditions, it is difficult to assess how well the approach generalizes to more diverse augmentation conditions.

  2. Limited consideration of image-dependent effects and experimental reliability. The proposed method considers augmentation conditions as the main factor for selection, but the effectiveness of augmentation may depend on both the condition and the source image. It is unclear whether selecting conditions globally is appropriate when the same condition may have different effects across images. Since the method averages similarities across source images and removes selected conditions from the entire dataset, such image-dependent differences are not considered or analyzed. In addition, the performance improvements are relatively small, with noticeable variation across training seeds. The experiments also use a relatively small dataset, with only 102 source images for measuring condition similarity and 247 images for external evaluation. This makes it difficult to assess the reliability of the reported improvements.

Questions for authors
  1. How does the proposed method compare with random condition selection? Is there a clear relationship between condition similarity and downstream performance gains?

  2. How does the method distinguish improvements from increased training data and those from greater condition diversity?

  3. How much do condition similarities vary across source images, and how might this affect the proposed global condition selection?

  4. How reliable are the reported improvements given the relatively small dataset?

3 comments
AuthorsOpenReview ↗

Additional Results and Manuscript Revisions (1/2)

Following up on our earlier response, we have completed the additional analyses and incorporated them into the revised manuscript. Below we highlight the new evidence addressing your concerns and indicate where each addition can be found.

Q1. Random selection and downstream performance

Section 5.3 and Table 4 now compare our method with uniform random selection, validation-AP-based selection, K-means, hierarchical clustering, and farthest-first selection. At each KK, methods share the generated-image count, clean/generated sampling proportion, training budget, and seeds. Random performance is calculated exactly over all same-size subsets. All selected subsets in Table 4 exceed their corresponding random expectation. At K=4K=4, our method also exceeds the clustering baselines by 1.26 AP for RT-DETRv2-S and 0.31 AP for YOLO11m. Appendix D provides the protocols and explains how redundancy distinguishes alternatives with equal coverage.

Section 5.4 and Figure 4 bring the response–transfer analysis into the main text. Across ten condition pairs, response similarity correlates with symmetric cross-condition transfer gains (Spearman ρ=0.879\rho=0.879; Pearson r=0.908r=0.908). Appendix C provides the matched duplicated-clean controls and per-seed results. This supports the task relevance of the measured relationships.

EvidenceSettingResult
Gain over clusteringRT-DETRv2-S, K=4K=4+1.26 AP
Gain over clusteringYOLO11m, K=4K=4+0.31 AP
Spearman correlation10 condition pairsρ=0.879\rho=0.879
Pearson correlation10 condition pairsr=0.908r=0.908

Q2. Training quantity diversity and condition count

Section 5.2 and Appendix B.2 isolate the generated sampling proportion from condition composition. For RT-DETRv2-S, SDL and matched Full use 6,000 successful optimizer updates and a generated sampling proportion of 3/83/8, while matched Full retains all five conditions. The original three-seed gain is 2.11 AP (Table 9); the extended 13-seed comparison gives a significant mean gain of 1.454 AP (Table 3). The same-K comparisons in Table 4 additionally hold training-pool size and condition count constant. Tables 1–2 and the complete subset results in Appendices A.3 and B.1 show performance across condition counts, including cases where adding conditions reduces AP.

RT-DETRv2-S comparisonValue
Successful optimizer updates, both methods6,000
Generated sampling proportion, both methods3/83/8
Conditions retained by matched Full5
SDL gain over matched Full, original 3 seeds+2.11 AP
SDL gain over matched Full, extended 13 seeds+1.454 AP
AuthorsOpenReview ↗

Additional Results and Manuscript Revisions (2/2)

Q3. Image dependence and global selection

Sections 3 and 4.2 specify the dataset-level selection objective and paired measurement: each condition is evaluated on the same training sources, similarities are computed within each source from generated-minus-clean responses, and these similarities are then averaged.

We now directly quantify image dependence for YOLO11m in Appendix A.2 and Figure 7. Across the 102 sources, the sample standard deviations of pairwise similarities range from 0.063 to 0.386; dust–low-light similarity is positive for all 102 sources. We further draw 10,000 bootstrap resamples of the sources, retaining all five variants of each sampled source together. For each resample, we recompute the response matrix and rerun selection at the fixed reported λ\lambda values.

Table 2 and Section 5.4 report recovery rates of 99.29% for RSFD, 100% for RSD, and 71.28% for RD. In the remaining RD resamples, snow is retained alongside rain and dust. These results quantify both the variation across images and its effect on the global decision, with particularly stable recovery of the three-condition RSD set.

Q4. Reliability with the available dataset

We strengthened the evaluation with ten additional randomly generated training seeds, giving 13 seeds including the original three. Section 5.2, Table 3, and Appendix B.3 report paired external AP differences, individual 95% t confidence intervals, and two-sided paired t-tests with Holm correction across the three comparisons:

ComparisonMean gain95% CIHolm-adjusted ppPositive runs
RT-DETRv2-S: SDL vs Full1.293[0.474, 2.111]0.011812/13
RT-DETRv2-S: SDL vs matched Full1.454[0.564, 2.344]0.011813/13
YOLO11m: RSD vs Full0.742[0.095, 1.390]0.028011/13

The gains are positive in 12/13, 13/13, and 11/13 paired runs, respectively; all three comparisons remain significant after correction. Table 10 lists the additional runs. Together with source-level resampling, these experiments strengthen evidence for repeatable downstream gains and selection stability within the available data. The external test set remains 247 images; the significance tests characterize training-run variability on this fixed benchmark.

Representativeness and extensibility of the candidate conditions

Section 3 now explains that rain, snow, fog, dust, and low light cover common degradations from precipitation, atmospheric obscuration, and reduced illumination. It also specifies how additional conditions enter the framework: generate paired variants on the same sources, extend the response matrix, and apply the same selection procedure.

We hope these additions help address your concerns about condition selection and experimental reliability.

Thank you to the authors for their additional experiments. The authors have addressed most of my concerns, including those regarding experimental reliability through the expanded seed evaluation. However, the limited test dataset size still remains a concern regarding generalizability.

Reviewer_DWieOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a condition-aware data augmentation framework for UAV detection under adverse environments. It proposes a two-stage generation pipeline that modifies image backgrounds while preserving UAV targets and their original bounding boxes, and selects training conditions through detector-guided feature analysis and a greedy coverage–redundancy strategy. Experiments on 730 Det-Fly images across five adverse conditions, using RT-DETRv2-S and YOLO11m with three random seeds, show that selected subsets of two to four conditions outperform both clean-only training and full-condition augmentation in mean external AP.

Critical strengths

The paper addresses a practical and well-motivated question of which adverse conditions should be generated for UAV detector training, and provides a clearly specified pipeline for condition selection and image generation. Experiments with RT-DETRv2-S and YOLO11m show that selected subsets of two to four conditions outperform both Clean-only and Full generation on mean external AP. The matched-ratio control further helps distinguish the benefits of condition selection from those of augmentation quantity, suggesting that selective adverse-condition generation could be a useful strategy for improving UAV detection robustness.

Critical weaknesses

The proposed condition selection strategy lacks comparisons with random subset selection and validation-AP-based selection. SDL improves over Full generation by only 0.80 AP, while another subset (RFD) achieves higher external AP (35.08 vs. 34.88). Internal validation favors RSDL, yet SDL is highlighted based on external performance, leaving the selection of λ unclear. Moreover, the experiments use only 730 clean images and 247 external test images, with substantial variation across three seeds and no statistical significance tests. The two-stage generation is also supported mainly by qualitative examples, without controlled downstream comparisons against single-stage generation or simple image compositing. These limitations weaken the evidence for the method's effectiveness and novelty.

Questions for authors

How is the trade-off parameter λ selected in practice? Since internal validation favors RSDL while SDL performs better on the external test set, was λ chosen independently of test performance?

Given the substantial variation across seeds in Table 5, is the reported 0.80 AP gain over Full generation statistically significant? Would this improvement remain reliable across additional seeds?

How does the proposed condition selection compare with random subset selection and validation-AP-based selection? Do the gains from two-stage generation persist when compared with simple image compositing and single-stage generation under the same training settings?

1 comment
AuthorsOpenReview ↗

Updated Manuscript: Additional Results and Clarifications

We have revised the manuscript to incorporate the following analyses and clarifications.

Q1. Selecting λ\lambda independently of test performance

Section 4.3 (pp. 5–6) now specifies a training-only rule: for a requested condition count, choose the distinct algorithm output with the highest coverage, break ties by lower redundancy, and take an interior λ\lambda from its stable interval. Section 5.4 (p. 10) reports the stable intervals for SDL and RSD and explains the default choice of SDL at K=3K=3. Section 5.1 and Table 1 report all distinct algorithm outputs, including RSDL and SDL; their internal and external AP rankings do not determine λ\lambda.

Appendix B.1, Table 7 (pp. 17–18), retains the complete subset results, including RFD's higher external AP. Our selection uses training-response coverage and redundancy; it does not optimize external test AP.

Q2. Sampling exposure and statistical reliability

Section 5.2 (pp. 7–8) and Appendix B.2, Table 9 (p. 18), explain the matched-sampling comparison. Full retains all five conditions while matching SDL's generated-image proportion of 3/83/8, the 6,000-update training budget, and final-EMA evaluation. Matching both training budget and generated-image exposure more directly isolates the contribution of condition selection. The original three-seed gain under this control is +2.11 AP.

Section 5.2, Table 3 (p. 8), now reports paired results over 13 seeds: the original three plus ten additional seeds. SDL gains +1.293 AP over ordinary Full (95% CI [0.474, 2.111]) and +1.454 AP over matched Full (95% CI [0.564, 2.344]); both have Holm-adjusted p=0.0118p=0.0118. Gains are positive in 12/13 and 13/13 seeds, respectively. The expanded ordinary-Full comparison supports the positive improvement originally reported as +0.80 AP. YOLO RSD also gains +0.742 AP over Full (Holm-adjusted p=0.0280p=0.0280). Appendix B.3, Table 10 (p. 19), lists the additional per-seed results and specifies the two-sided paired t-tests, confidence intervals, and Holm correction across the three comparisons.

ComparisonSeedsGain (AP)95% CIHolm-adjusted ppPositive seeds
RT SDL vs. ordinary Full3+0.80———
RT SDL vs. matched Full3+2.11———
RT SDL vs. ordinary Full13+1.293[0.474, 2.111]0.011812/13
RT SDL vs. matched Full13+1.454[0.564, 2.344]0.011813/13
YOLO RSD vs. Full13+0.742—0.0280—

—: not specified in this reply.

Q3a. Random and validation based selection baselines

Section 5.3, Table 4 (pp. 8–9), adds uniform random selection, validation-AP-based selection, K-means, average-linkage hierarchical clustering, and farthest-first selection. Comparisons match condition count, generated-image count, sampling ratio, training budget, and training seeds. Appendix D.1–D.2 and Table 16 (p. 23) specify the procedures, including the held-out internal validation criterion and tie handling. All eight outputs in Table 4 exceed their same-size random expectations; six outperform validation-based selection and two select the same subset.

Q3b. Single stage generation and simple compositing

Section 5.5, Figure 5 (p. 10), and Appendix E, Figure 9 (pp. 25–26), expand the evidence on target preservation. The nine illustrated single-stage FLUX.2-dev trial outputs include target displacement, enlargement beyond the source boxes, and additional unannotated UAVs. These annotation-breaking failures would introduce incorrect supervision if the original labels were reused. We therefore did not extend this pilot to large-scale downstream training with unchanged source annotations; the failures motivated separating background editing from source-target restoration.

Our central question is which conditions to generate and combine for detector training, as clarified in the Introduction and Section 3. However, we study this question within a fixed diffusion-based generation pipeline, shared by Full and the condition-selection baselines. A matched-training comparison with simple compositing would evaluate an alternative generation method. The reported AP gains establish the benefit of condition selection within the shared pipeline.

Reviewer_TgEEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a detector-guided generative data augmentation framework for improving UAV detection under adverse environmental conditions. The authors address the problem of selecting useful augmentation conditions, motivated by the observation that semantically different conditions may produce similar changes in detector feature representations. The proposed framework consists of two main components: a two-stage decoupled generation process that modifies the background and reintegrates the original UAV using soft-mask blending to preserve object annotations, and a detector-guided condition-selection strategy that measures context-feature changes using a frozen detector trained on clean images. Based on pairwise cosine similarities of these feature changes, the authors formulate a coverage-redundancy objective and employ a backward greedy algorithm to select a subset of augmentation conditions. The selected synthetic images are combined with clean training data to train downstream detectors. Experiments involving five adverse conditions (rain, snow, fog, dust, and low light) and two detection architectures (RT-DETRv2-S and YOLO11m) show that the selected two- to four-condition subsets achieve higher external mean average precision (AP) than both Clean-only training and Full generation under the evaluated protocols. The authors conclude that selecting augmentation conditions based on detector responses can improve UAV detection performance compared with indiscriminately using all available generated conditions.

Critical strengths

The paper addresses an important problem of choosing useful weather conditions for improving UAV detection instead of using all generated images. The authors use changes in detector features to measure how different weather conditions affect detection. The two-stage generation method also helps preserve the original UAV and its annotations while changing the background. This is a practical approach that could help select more useful synthetic training data.

The authors evaluate their method using two object detectors, RT-DETRv2-S and YOLO11m, and five weather conditions. They test different combinations of conditions and repeat the experiments using three random seeds. They also compare training with a matched ratio of clean and generated images for RT-DETRv2-S. The selected conditions improve external mean AP over Full generation by 0.80 points for RT-DETRv2-S and up to 1.14 points for YOLO11m. The supplementary material includes source code and experimental details, which support reproducibility. These results provide useful evidence that selecting appropriate augmentation conditions can improve overall detection performance.

Critical weaknesses

One major weakness is that the reported improvements do not clearly show better performance under real-world adverse weather conditions. The authors calculate external mean AP by averaging results from 130 real-world images and 117 synthetic images. However, the improvements mainly come from the synthetic test set. For RT-DETRv2-S, the selected Snow/Dust/Low-light subset achieves 38.31 AP on the real-world test set, compared with 38.28 AP for Clean-only training. This is an improvement of only 0.03 AP, although the overall external AP improves by 2.95 points. Similarly, for YOLO11m, the selected Rain/Snow/Dust subset achieves 31.06 AP on real-world images, which is lower than the 32.69 AP obtained with Clean-only training. These results raise concerns about whether the proposed method can consistently improve UAV detection in real-world adverse weather conditions. The authors should discuss these differences and provide a clearer analysis of real-world and synthetic test performance.

Another important weakness is that the paper does not clearly demonstrate why the proposed condition-selection method is better than simpler approaches. The authors use detector feature similarities and a backward greedy algorithm to choose useful augmentation conditions, but they do not compare their method with alternatives such as random selection or clustering-based selection. Although the selected subsets generally perform well, they do not always achieve the highest detection accuracy among the tested combinations. In addition, since only five conditions are considered, all possible subsets can be evaluated to find the best solution for the proposed selection objective. The greedy algorithm does not always find this solution, and a higher objective value does not necessarily lead to better detection performance. Therefore, more evidence is needed to show that the proposed selection method is effective and provides a clear advantage over simpler alternatives. This is particularly important because condition selection is the main contribution of the paper.

Questions for authors

Q1. The external mean AP combines results from real-world adverse-weather images and synthetic UAV images. However, the supplementary results show that most of the improvement over Clean-only comes from the synthetic test set, while the improvement on real-world images is very small or negative. Could the authors explain why the method shows limited improvement on real-world adverse-weather images? Could they also report the average AP and variation across different runs for both test sets separately?

Q2. The paper proposes a detector-guided method to select useful augmentation conditions, but it does not directly compare this method with simpler approaches, such as random selection or clustering-based selection. Could the authors provide these comparisons using the same number of selected conditions? This would help clarify whether the performance improvements come from the proposed selection method or simply from removing less useful augmentation conditions.

Q3. The paper uses a backward greedy algorithm to select augmentation conditions. Since only five conditions are considered, it is possible to examine all combinations and find the subset that maximizes the proposed objective. However, the greedy algorithm does not always select this subset, and a higher objective value does not necessarily produce better detection performance. Could the authors explain why backward greedy selection was chosen and how well the proposed objective relates to actual detection accuracy? Also, how should the parameter λ be selected for a new dataset without using test results? The paper addresses an important problem in UAV detection and proposes an interesting method that combines two-stage image generation with detector-guided condition selection. The experiments using two detectors, different combinations of weather conditions, and multiple training runs are useful strengths. The supplementary material also provides code and experimental details.

However, I have two main concerns. First, most of the reported improvement over Clean-only training comes from the synthetic test set, while performance on real-world adverse-weather images shows little improvement or even decreases. This raises questions about how well the method improves detection under real-world conditions. Second, the paper does not provide enough evidence that the proposed condition-selection method is better than simpler approaches, such as random selection or clustering-based selection. Since condition selection is the main contribution, this comparison is important.

For these reasons, I recommend Weak Rejection in the current form. However, I would reconsider my assessment if the authors provide a clear explanation of the real-world evaluation results and stronger evidence that their selection method performs better than simpler alternatives.

Comments on the paper

1 comment

Dear reviewers,

As the discussion period is limited, please follow up on the latest responses in a timely manner to address any remaining concerns. Thank you!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.