acceptodds
Under review as a conference paper at ICLR 2027

GOLD: Generalized Proposal Denoising for Domain-Generalized Object Detection

Abstract

Domain-generalized object detection (DGOD) aims to train a detector solely on labeled source domain data that generalizes directly to unseen target domains without accessing target data or performing test-time adaptation. Despite encouraging progress, existing approaches largely optimize encoded features or final predictions within deterministic detection frameworks, and thus lacking an explicit mechanism to iteratively recover object locations from unreliable spatial hypotheses. To this end, we propose **G**eneralized Pr**O**posa**L** **D**enoising **(GOLD)**, a diffusion-based DGOD framework that reformulates cross-domain localization as a semantics-preserving denoising process that progressively refines noisy proposals under unreliable visual evidence. First, a Spatial Calibration Adapter (SCAdapter) equips frozen DINOv3 features with localization-sensitive spatial structure for iterative proposal refinement. Second, Spatially Heterogeneous Evidence Perturbing (SHEP) decomposes visual evidence into stochastic foreground regions, localization-critical contexts, and distant backgrounds, as well as applies moderate, strong, and weak residual perturbations, respectively, according to their roles in proposal denoising. Third, Prototype-Anchored Denoising (PAD) constructs multi-scale category prototypes and anchors matched proposal representations at every refinement stage, preserving category semantics throughout proposal denoising. Extensive experiments on multiple cross-domain benchmarks demonstrate that GOLD consistently outperforms previous DGOD methods. *Our code is released in supplementary materials.*

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Domain Generalization, Object Detection, Proposal Denoising

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KGcUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose GOLD, a diffusion-based framework for domain-generalized object detection that formulates cross-domain localization as semantics-preserving proposal denoising. Built on DiffusionDet, GOLD combines SCAdapter for spatially calibrated DINOv3 features, SHEP for region-dependent visual perturbations, and PAD for prototype-guided proposal refinement. Experiments on Diverse Weather, Cityscapes→BDD100K, and Cityscapes→FoggyCityscapes demonstrate consistent improvements over existing DGOD methods.

Critical strengths
  1. The paper proposes GOLD, a diffusion-based DGOD framework built on DiffusionDet. It combines SCAdapter for spatially adapted DINOv3 features, SHEP for region-specific perturbations, and PAD for multi-scale prototype guidance during proposal refinement.
  2. The method is evaluated on Diverse Weather, Cityscapes→BDD100K, and Cityscapes→Foggy Cityscapes. Ablation studies examine the contributions of SCAdapter, SHEP, and PAD, as well as perturbation strengths and prototype design choices. The paper also reports training overhead and inference cost.
  3. The method and optimization objectives are clearly described, with pseudocode and hyperparameter tables provided. Stage-wise box refinement, Grad-CAM, and t-SNE visualizations further support the analysis. The work focuses on maintaining reliable spatial evidence during iterative localization under domain shifts.
Critical weaknesses
  1. SCAdapter is motivated by the claim that frozen DINOv3 features exhibit a mismatch between semantic representation and object localization. However, the paper does not provide direct experimental evidence to verify this claim or isolate the localization-specific contribution of SCAdapter.
  2. SHEP applies several perturbations to the input images, but real-world visual perturbations are diverse and difficult to exhaustively cover during training, including snow, spatter, frost, shot noise, impulse noise, speckle noise, motion blur, JPEG compression, pixelation, and geometric deformation. The paper does not evaluate GOLD on such unseen perturbations, leaving its robustness under out-of-distribution degradations unclear.
  3. The scalability of GOLD beyond the frozen DINOv3 backbone has not been demonstrated. Without experiments using alternative visual encoders or architectures, it remains unclear whether the reported improvements are generally applicable or primarily dependent on DINOv3’s pretrained representations.
Questions for authors
  1. SCAdapter is motivated by the claim that frozen DINOv3 features exhibit a mismatch between semantic representation and object localization. Is there experimental evidence to support this claim?
  2. SHEP applies several perturbations to the input images, but real-world visual perturbations are diverse and difficult to exhaustively cover during training, including unseen cases such as snow, spatter, frost, shot noise, impulse noise, speckle noise, motion blur, JPEG compression, pixelation, and geometric deformation. How well does the proposed method perform when directly applied to test images with these unseen perturbations? This would provide a more intuitive evaluation of GOLD’s actual contribution.
  3. The scalability of the method to other architectures has not yet been demonstrated. If DINOv3 is replaced with other visual encoders, how much performance improvement can the method still achieve?
5 comments
AuthorsOpenReview ↗

Response to Reviewer KGcU Weakness 1 & Question 1

KGcU - Weakness 1 & Question 1

A. Clarification of SCAdapter's Motivation.

Thank you for encouraging us to clarify the motivation and supporting evidence. We will revise the wording in lines 169–172, with the changes highlighted in green, to emphasize task-specific spatial adaptation rather than a general limitation of DINOv3 representations:

Pretrained features benefit from further task-specific adaptation to support spatially sensitive proposal refinement. We therefore introduce a trainable Spatial Calibration Adapter (SCAdapter) at each feature level to incorporate local spatial information before multi-scale fusion.

We have also revised the corresponding descriptions in the second contribution and the Conclusion to maintain consistent wording throughout the manuscript. These changes are highlighted in green.

B. Controlled Ablations of SCAdapter.

Importantly, Table 9 (rows 1 and 5) already provides a controlled comparison: with the same DINOv3 backbone and SHEP/PAD enabled, replacing linear adaptation with SCAdapter increases the average mAP from 53.5% to 56.8% (+3.3% mAP). We further isolate SCAdapter without SHEP/PAD in Table KGcU 1: rows 1 and 2 use the same frozen DINOv3 backbone, and SCAdapter improves the average from 46.0% to 48.8% (+2.8% mAP). All of these results can prove that SCAdapter helps improve the positioning capability of the model.

Table KGcU 1. Comparison of linear adaptation and SCAdapter on Diverse Weather, without SHEP or PAD.

AdapterDFDRNRNSAvg.
Linear Adapter44.652.335.651.646.0
SCAdapter48.352.042.752.148.8
AuthorsOpenReview ↗

Response to Reviewer KGcU Weakness 2 & Question 2

KGcU - Weakness 2 & Question 2

The public imagecorruptions implementation provides standardized noise, blur, weather, and digital corruptions, including snow, spatter, frost, shot noise, impulse noise, speckle noise, motion blur, JPEG compression, and pixelation. We are conducting an additional evaluation on Cityscapes-Corruption using models trained on Cityscapes:

Table KGcU 2. Generalization results on Cityscapes-Corruption. All models are trained on Cityscapes. For each type of corruption, the average mAP is reported over five severity levels, and mPC denotes the mean value over all types.

MethodVenuesNoiseBlurWeatherDigitalmPC
GaussShotImpulseDefocusGlassMotionZoomSnowFrostFogBrightContrastElasticPixelJPEG'
StylizedAugICLR'194.86.84.319.518.718.52.73.517.030.531.922.733.922.620.817.2
AutoAugECCV'200.91.60.916.814.418.92.01.916.032.935.226.339.417.911.615.8
AugMixICLR'205.06.85.118.318.119.36.25.020.531.233.725.637.420.319.618.1
SupConNeurIPS'207.09.57.422.620.222.34.35.323.037.338.931.640.124.020.120.9
FSCECVPR'217.410.28.223.320.321.54.85.623.637.138.031.940.023.220.421.0
OA-MixAAAI'247.29.67.722.818.821.95.45.223.637.338.731.940.222.220.220.8
OA-DGAAAI'248.210.68.424.620.522.34.86.125.038.439.732.840.223.822.021.8
PhysAugAAAI'2514.317.011.925.619.125.53.98.621.335.339.527.539.128.919.922.6
OPCDMM'2518.020.113.725.120.524.94.013.222.839.039.131.937.326.220.123.7
UICLAAAI'2628.530.929.231.921.626.014.613.637.244.540.538.241.638.525.831.5
Ours-30.832.632.033.824.027.616.816.239.047.042.040.943.740.428.133.0

As shown in Table KGcU 2, GOLD achieves the highest mPC of 33.0%, outperforming the strongest baseline (UICL) by 1.5% mAP. Consistent improvements are observed across all 15 corruption types, spanning noise, blur, weather, and digital degradations. These results demonstrate that GOLD maintains strong robustness against diverse out-of-distribution perturbations, including those not explicitly considered during training. We also incorporate these results into the experiment section of the revision as Table 12 to further demonstrate GOLD's robustness against unseen visual perturbations (highlighted in green).

AuthorsOpenReview ↗

Response to Reviewer KGcU Weakness 3 & Question 3

KGcU - Weakness 3 & Question 3

We evaluate GOLD with ResNet-50, ResNet-101, Swin-B, and DINOv3 in DiffusionDet, matching the pretrained checkpoint and training settings within each baseline–GOLD pair. Table KGcU 3 reports both per-domain performance and the gain in four-domain average mAP.

Table KGcU 3. Cross-encoder evaluation on Diverse Weather.

Visual encoderMethodDFDRNRNSAvg.Gain
ResNet-50Baseline36.529.015.736.829.5—
ResNet-50GOLD43.043.126.946.840.0+10.5
ResNet-101Baseline38.130.817.338.231.1—
ResNet-101GOLD43.743.928.447.740.9+9.8
Swin-BBaseline42.349.228.751.242.9—
Swin-BGOLD51.557.244.455.752.2+9.3
DINOv3Baseline44.652.335.651.646.0—
DINOv3GOLD55.662.348.860.556.8+10.8

Replacing DINOv3 with ResNet-50, ResNet-101, or Swin-B still yields 10.5%, 9.8%, and 9.3% mAP gains, respectively, with improvements on every target domain. DINOv3 gains 10.8 points and achieves the highest final average mAP (56.8%). These results show that GOLD's benefits extend across the evaluated CNN and Transformer encoders, while encoder choice continues to affect the final performance. We have added these results and a detailed analysis to Appendix F, highlighted in green.

Reviewer_mbJUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GOLD, a framework for domain-generalized object detection built on DiffusionDet. It aims to improve the robustness of iterative proposal refinement under domain shifts by adapting pretrained visual features and incorporating spatially heterogeneous augmentation and prototype-based semantic guidance. The authors evaluate the framework on several cross-domain detection benchmarks and report improvements over existing methods.

Critical strengths

S1. Clear and well-organized presentation.

The paper is clearly written and easy to follow. Its well-organized structure, clear illustrations, and clean layout help communicate the proposed framework and its components.

S2. Interesting augmentation strategy.

The proposed spatially heterogeneous augmentation strategy (SHEP) is interesting and intuitive. Applying different perturbation strengths to foreground, localization-related context, and background regions accounts for their distinct roles in detection and offers a sensible approach to improving robustness under domain shift.


However, the key concerns outlined in the Weaknesses section substantially affect my overall assessment and rating of the paper.

Critical weaknesses

W1. The diffusion-specific contribution is insufficiently established.

The paper attributes improved generalization to semantics-preserving proposal denoising, yet its three modules primarily involve feature adaptation, region-wise augmentation, and prototype supervision. Their specific connection to diffusion modeling remains unclear, and iterative box refinement is not unique to diffusion-based detectors. While the ablations support the modules’ effectiveness, they do not distinguish the benefits of generic multi-stage training from those of diffusion modeling. A controlled comparison with a non-diffusion iterative detector under matched settings is missing.

W2. Backbone priors and methodological contributions are not sufficiently disentangled.

GOLD uses a frozen DINOv3 ConvNeXt-L backbone, but the main tables do not specify the backbones and pretraining configurations of competing methods, and the DiffusionDet baseline configuration remains unclear. Table 8 does not clearly separate the contribution of introducing DINOv3 from that of SCAdapter. Moreover, detection accuracy for a matched-backbone baseline without the proposed modules is not clearly reported, making it difficult to determine how much of the overall gain comes from pretrained representations versus the proposed design.

Questions for authors

Please refer to W1 and W2 above for the context of the following questions.

Q1. Non-diffusion detectors such as Cascade R-CNN also progressively refine proposals using visual evidence. Why does unreliable visual evidence under domain shift specifically motivate the use of DiffusionDet? Please clarify what diffusion modeling contributes beyond conventional multi-stage refinement.

Q2. Could the authors evaluate the proposed modules within a common detection framework, such as Faster R-CNN, and compare against existing DGOD methods using the same backbone and pretraining, with comparable training settings? This would allow a more controlled assessment of the proposed method’s improvements by reducing differences in detector architectures and pretrained representations.

Q3. Could the authors specify the backbone and pretraining configurations of the compared methods, particularly the DiffusionDet baseline, and clarify how differences in these configurations are accounted for to ensure a fair comparison?

9 comments
AuthorsOpenReview ↗

Response to Reviewer mbJU Weakness 1

A. Connection to proposal denoising.

Our use of diffusion refers specifically to DiffusionDet's recovery of object boxes from noisy box coordinates, rather than image synthesis or feature-space diffusion. As described in lines 62–71, our three modules are closely integrated with this box-denoising process, enhancing its conditioning features, robustness to visual perturbations, and proposal representations throughout iterative refinement.

B. Distinction from conventional multi-stage refinement.

We agree that iterative bounding-box refinement is not exclusive to diffusion-based detectors, as demonstrated by methods such as Cascade R-CNN. Our intention is not to claim that DiffusionDet uniquely performs iterative refinement, but rather to emphasize its distinct noise-conditioned box-recovery formulation.

To avoid an overly general statement, we have revised "these methods" to "most existing methods" in lines 48–49 of the Introduction, with the changes highlighted in red. Furthermore, Lines 52–54 describe DiffusionDet as offering "a different perspective", rather than claiming that iterative refinement is exclusive to diffusion-based detection.

Unlike conventional cascade refinement, DiffusionDet learns to recover object boxes from proposals degraded according to a diffusion noise schedule, with explicit diffusion-time conditioning. Our motivation is therefore to investigate how this particular recovery process can remain reliable when the visual representations guiding it are affected by domain shift. The following table further clarifies this distinction:

Table mbJU 1. Comparison of Cascade R-CNN and DiffusionDet.

AspectCascade R-CNNDiffusionDet
Multi-stage box refinementYesYes
Proposal initialization at inferenceRPN-generated proposalsRandomly sampled boxes
Training proposalsRPN proposals and outputs of preceding cascade stagesGround-truth boxes corrupted according to a diffusion noise schedule
Refinement formulationCascaded classification and box regression with increasing IoU thresholdsNoise-conditioned recovery of object boxes
Explicit diffusion-time conditioningNoYes

DiffusionDet's defining mechanism is the recovery of object boxes from noisy boxes. GOLD asks how to maintain localization accuracy and semantic stability when the visual evidence guiding this recovery is affected by domain shift. We provide controlled comparisons with non-diffusion iterative detectors (Cascade R-CNN, Sparse R-CNN) in our response to your Question 1.

AuthorsOpenReview ↗

Response to Reviewer mbJU Weakness 2 [Part 1/2]

A. Backbone and pretraining transparency.

We will add a Backbone column to Table 1, with the additions highlighted in red. This will make the representation differences between the compared methods explicit.

Table mbJU 2. (Revised Table 1) Results on Diverse Weather with backbone information.

MethodVenueBackboneDFDRNRNSAvg.
S-DGODCVPR’22ResNet-10133.528.216.636.628.7
CLIP the GapCVPR’23ResNet-10138.532.318.736.931.6
SRCDTNNLS’24ResNet-10135.928.817.036.729.6
PDOCCVPR’24ResNet-10139.133.719.238.532.6
UFRCVPR’24ResNet-10139.633.219.240.833.2
G-NASAAAI’24ResNet-10136.435.117.445.033.5
DivAlignCVPR’24ResNet-10137.238.124.142.535.5
FWCLTMM’25ResNet-10132.332.618.937.530.3
GDDCVPR’25Stable Diffusion 1.543.342.527.847.040.2
BoostICCV’25Stable Diffusion 1.545.944.328.248.841.8
SE-COTCVPR’25Swin-T44.949.533.752.745.2
DoSLCVPR’26ResNet-10141.240.624.844.037.7
CD-FKDICRA’26ResNet-101-FPN40.242.323.447.338.3
UICLAAAI’26ResNet-101-FPN40.142.227.445.238.7
LTFEAAAI’26Swin-T53.146.533.446.444.9
VFM4 SDGarXiv’26DINOv3 ViT-L/1647.450.729.056.145.8
MR-DCoTTPAMI’26Swin-T48.353.236.454.648.1
BridgeCVPR’26DINOv3 ViT-L/1648.153.846.754.650.8
DiffusionDetICCV’23Swin-B42.349.228.751.242.9
GOLD (Ours)—DINOv3 ConvNeXt-L55.662.348.860.556.8

B. Separating backbone adaptation from SCAdapter.

Table 8 presents an overall ablation study of the three proposed modules, while Table 9 provides a detailed ablation of SCAdapter (backbone). In our implementation, the linear adapter represents the minimal architectural modification required to integrate DINOv3 into the detection framework, serving as a baseline to isolate the contribution of SCAdapter. Table 9 already compares linear adaptation and SCAdapter with the same frozen DINOv3 ConvNeXt-L backbone and with SHEP/PAD enabled: rows 1 and 5 achieve 53.5% and 56.8%, respectively, demonstrating a +3.3% improvement attributable to SCAdapter.

We further isolate SCAdapter without SHEP or PAD in Table mbJU 3: replacing linear adaptation with SCAdapter increases the average mAP from 46.0% to 48.8% (+2.8%). Both configurations use the same pretrained backbone, demonstrating the contribution of SCAdapter beyond the backbone itself.

Table mbJU 3. Comparison of linear adaptation and SCAdapter on Diverse Weather, without SHEP or PAD.

AdapterDFDRNRNSAvg.
Linear Adapter44.652.335.651.646.0
SCAdapter48.352.042.752.148.8
AuthorsOpenReview ↗

Response to Reviewer mbJU Weakness 2 [Part 2/2]

C. Comparisons with matched backbone variants.

To separate the effects of pretrained representations from our proposed modules, Tables mbJU 4–5 compare DiffusionDet (Baseline) and GOLD using the same DINOv3 backbone at each scale, with all models trained on Cityscapes. We report per-class AP and mAP on BDD100K and Foggy Cityscapes separately. GOLD improves mAP by 12.1–13.3% on BDD100K and 12.5–13.7% on Foggy Cityscapes across the four backbone scales. It also improves every category over the corresponding baseline in both datasets. These matched-backbone results support the contribution of our proposed modules beyond the choice of pretrained backbone. We also add this new experiment results in Appendix D.1 (highlighted in red).

Table mbJU 4. Per-class detection results (%) on BDD100K for DiffusionDet (Baseline) and GOLD with matched DINOv3 backbones. All models are trained on Cityscapes.

MethodBusBikeCarMotorPersonRiderTruckmAP
DINOv3-T (Baseline)44.729.357.535.645.839.745.642.6
DINOv3-T (GOLD)55.641.374.852.362.854.050.555.9
DINOv3-S (Baseline)46.531.559.937.648.142.147.944.8
DINOv3-S (GOLD)55.942.876.755.367.354.950.357.6
DINOv3-B (Baseline)47.932.663.040.648.243.748.146.3
DINOv3-B (GOLD)57.842.778.456.469.555.450.558.7
DINOv3-L (Baseline)48.930.466.742.950.643.948.547.4
DINOv3-L (GOLD)58.343.679.157.569.956.751.259.5

Table mbJU 5. Per-class detection results (%) on Foggy Cityscapes for DiffusionDet (Baseline) and GOLD with matched DINOv3 backbones. All models are trained on Cityscapes.

MethodBusBikeCarMotorPersonRiderTrainTruckmAP
DINOv3-T (Baseline)60.941.059.636.150.750.955.133.748.5
DINOv3-T (GOLD)73.856.172.752.663.664.967.046.962.2
DINOv3-S (Baseline)64.543.161.739.453.653.257.534.350.9
DINOv3-S (GOLD)72.957.973.253.563.766.971.451.563.9
DINOv3-B (Baseline)63.545.462.842.552.552.860.439.352.4
DINOv3-B (GOLD)74.459.674.656.064.667.872.051.865.1
DINOv3-L (Baseline)63.546.263.041.454.855.660.639.553.1
DINOv3-L (GOLD)74.160.074.755.465.168.773.453.265.6
Reviewer_9WTKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GOLD, a single-source domain-generalized object detector built on DiffusionDet with a frozen DINOv3 ConvNeXt-L backbone and three training-time additions: a 1x1/DW3x3/1x1 SCAdapter on each encoder level (Eq. 2); SHEP, a photometric augmentation blended with region-dependent strengths $\rho_{bg} < \rho_{fg} < \rho_{ctx} over GT-derived foreground-support/context/background masks (Eqs. 3-8); and PAD, an InfoNCE-style loss pulling matched proposal features toward EMA class prototypes pooled from clean DINOv3 features at all six refinement stages (Eqs. 9-13). SHEP and PAD are dropped at inference. It reports 56.8 average [email protected] on Diverse Weather (Table 1, +6.0 over Bridge), 59.5 on Cityscapes ->BDD100K (Table 6) and 65.6 on Cityscapes -> Foggy Cityscapes (Table 7), with single-run ablations (Tables 8-11), a cost study (Table 15), and supplied code.

Critical strengths
  • Large, consistent headline numbers. +6.0 average mAP over Bridge (50.8) on Diverse Weather, leading all four domains (Night-Rainy 28.7 -> 48.8), plus 59.5 on Cityscapes -> BDD100K and 65.6 on Foggy Cityscapes; if the gains survive backbone-matched controls, the result would be strong.
  • Internally consistent ablations, precise SHEP, supplied code, honest cost disclosure. Tables 8-11 share rows (52.7, 53.6, 56.8) and every average recomputes; Eqs. (3)-(8) match shep.py:100-170; Table 15 shows identical inference cost and 29.8-38.3% lower training throughput (l. 751-752). This made a full code audit possible and makes the paper fixable in principle.
Critical weaknesses
  • No variance, and a target-set evaluation loop. No variance, seed count, or number of runs is reported, although inference is stochastic (gold.py:248, fixed seed yaml:136), and conclusions rest on 0.5-1.0 mAP gaps (Table 11 rows 3 vs 5 +0.5; Table 10 rows 9 vs 10 +1.0; Table 6 +0.6); the four target test sets are evaluated every 10k iterations (yaml:94,133) with no source validation split or selection procedure described (DW_Daytime_Clear_test exists in builtin.py) while ρ ranges, d_min = 24 px, m_min = 12 px, and ω\omega look finely tuned. Until seeds and selection are reported, the sub-1-mAP conclusions cannot be separated from noise, nor the headline numbers from test-domain selection.
  • Code and paper disagree, with a provenance question. Paper l. 236-237 (jitter + Gaussian blur + Gaussian noise) vs code shep.py:213-218/config.py:78-86 (jitter + gamma + motion/Gaussian blur, p = 0.26, no noise); shep.py:38-40 mentions removed "non-paper weather operators", raising the question of whether Tables 1-11 were produced with weather-synthesis augmentation that would inject target-domain knowledge. Until clarified, it is not established that the released code reproduces the reported numbers or that they were obtained without target-weather knowledge.
Questions for authors
  1. Does Table 8 row 2 (DINOv3+SCAdapter, no SHEP/PAD, 48.8) use the same resolution, proposals, and schedule as row 5; what is the corresponding accuracy under linear adaptation with no SHEP/PAD; what are the DINOv3 baseline accuracies on Cityscapes -> BDD100K and Foggy Cityscapes (the Table 15 models); what settings does the "DiffusionDet (ICCV'23)" row use; and were the other Table 1-7 numbers re-run or taken from the original papers?
  2. Were Tables 1-11 produced with or without the weather-synthesis operators mentioned in shep.py:38-40, have they been reproduced with the released operator set, and which ω\omega vector was used (paper/yaml vs config.py:102)?
  3. Are results from the final 450k checkpoint or the best periodic target-domain evaluation (yaml:94,133), and on what data were ρ ranges, ω\omega, dmind_{min}, mminm_{min} tuned?
  4. The paper states β=0.02\beta = 0.02 for GOLD (l. 320-321): which Foggy Cityscapes split and AP interpolation (COCO 101-point vs VOC) were used for each copied Table 7 baseline, and how many DDIM steps were used for GOLD?
  5. Why are Jiang et al. (ICCV'25), Cauvis (NeurIPS'25), and OA-DG (AAAI'24) absent, and can Bridge/LTFE per-class numbers be added to Tables 2-5?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 9WTK Weakness1 & 2

9WTK — Weakness 1

A. Reproducibility and variability.

Thank you for highlighting the need to quantify variability. Our original tables presented point estimates, as do the ablation tables in prior DGOD works. We agree that repeated evaluations provide a more informative assessment of reproducibility.

As shown in line 136 of the code configs/gold_diverse_weather.yaml, we are using a random seed. We additionally report five repetitions (five different random seeds) on the Diverse Weather Datasets:

Table 9WTK 1. Results of five repeated runs on Diverse Weather.

RunDFDRNRNSAvg.Best Iteration
155.562.047.959.856.3120k
255.362.548.661.457.0140k
356.161.948.461.156.9120k
455.863.547.160.556.7150k
555.961.849.560.356.9140k

The average ranges from 56.3 to 57.0, a spread of 0.7% mAP. The originally reported average of 56.8 is close to the repeated-experiment mean of 56.7. These results support the stability of the aggregate performance under the evaluated repetition protocol.

B. Interpretation of the ablation gains.

The proposed components yield substantial improvements: in Table 8, adding SHEP and PAD to the configuration with SCAdapter increases the average mAP from 48.8% to 56.8% (+8.0%). Adding PAD to SCAdapter + SHEP yields a gain of 4.1%, while adding SHEP to SCAdapter + PAD yields a gain of 3.2%. Tables 9–11 further investigate fine-grained design choices for SCA, SHEP, and PAD, respectively, examining how each component should be configured rather than solely whether it should be included. The smaller differences between these variants therefore reflect the incremental benefits of specific design choices within an already effective component. They should be distinguished from the larger component-level gains demonstrated in Table 8.

Similar incremental improvements can also be observed in the ablation studies of existing DGOD methods. For example, in Bridge, the proposed Sample Queries module brings improvements of 0.5%–0.7%, while in Unbiased Faster R-CNN, incorporating additional constraints yields gains of only 0.2%–0.7%. These results suggest that relatively small improvements are not uncommon in fine-grained ablation studies of DGOD methods.

Bridge (CVPR26), Table 6. Gain is relative to the preceding row.

LRBSQDINOv3SDSAMGain (DINOv3 / SD / SAM)
——57.751.845.8—
✓—60.953.149.2+3.2 / +1.3 / +3.4
✓✓61.653.649.9+0.7 / +0.5 / +0.7

Unbiased Faster R-CNN (CVPR24), Table 6.

RowMethodLatt\mathcal L_{att}Lexp\mathcal L_{exp}Limp\mathcal L_{imp}D-ClearN-ClearD-FoggyAvg.Gain
U1FR———56.235.833.541.833—
U2FR + GLT———60.438.636.145.033+3.200 vs. U1
U3FR + GLT✓——59.539.738.145.767+0.733 vs. U2
U4FR + GLT✓✓—59.140.439.046.167+0.400 vs. U3
U5FR + GLT✓—✓59.340.038.545.933+0.167 vs. U3
U6FR + GLT✓✓✓58.640.839.646.333+0.167 vs. U4

C. Hyperparameter specification and selection.

We follow the standard dataset splits used in prior DGOD work. The hyperparameters were selected through preliminary experiments to implement our spatial perturbation and stage-dependent supervision principles. To maintain a clear focus on the methodological contributions, we did not elaborate on every hyperparameter choice in the original manuscript.

9WTK - Weakness 2

Thank you for identifying an inaccurate description of the augmentation operators. We will correct it to “color jittering, gamma adjustment, and Gaussian blurring,” with the revision highlighted in blue. This matches the provide code diffusiondet/shep.py.

The implementation in diffusiondet/shep.py directly corresponds to the SHEP formulation in Section 4.3. These operations introduce generic appearance variations; they do not explicitly synthesize fog or other specific weather conditions. This design serves our objective of generalizing to diverse, unknown domains. We provided the implementation code to make these operations transparent and reproducible, and welcome all reviewers to inspect the correspondence between the formulation and the code.

AuthorsOpenReview ↗

Response to Reviewer 9WTK Question 1

9WTK - Question 1

A. Consistent experimental settings.

All experiments conducted by us use same image resolutions, proposal settings, and training schedules within each benchmark. The component ablations in Table 8 keep all settings unchanged except for the components being investigated.

B. Additional linear-adapter baseline.

in Table 9WTK 2 row 2, we add a new baseline using the same frozen DINOv3 backbone with a linear adapter, without SHEP or PAD. Replacing the linear adapter with our proposed SCAdapter increases the average mAP from 46.0% to 48.8%, yielding a gain of 2.8%.

Table 9WTK 2. Updated component ablation of GOLD on the four Diverse Weather target domains.

RowLinear AdapterSCAdapterSHEPPADDFDRNRNSAvg.
142.349.228.751.242.9
2✓44.652.335.651.646.0
3✓48.352.042.752.148.8
4✓✓52.258.144.555.952.7
5✓✓53.258.645.357.153.6
6✓✓✓55.662.348.860.556.8

C. Generalization performance with different DINOv3 backbones.

We have conducted additional experiments to report the BDD100K and Foggy Cityscapes performance of the baseline and GOLD configurations listed in Table 16. We have accordingly updated Table 16 in the Appendix D.1, with the revisions highlighted in blue. The updated table in the appendix now presents both computational efficiency and cross-domain generalization performance with matched DINOv3 backbones.

Table 9WTK 3. Training and inference efficiency with matched DINOv3 backbones, together with cross-domain generalization performance on BDD100K and Foggy Cityscapes.

MethodTraining throughput (it/s)Training GPU memory (GB)Inference throughput (img/s)Inference GPU memory (GB)BDD100K (mAP)Foggy Cityscapes (mAP)
DINOv3-T (Baseline: DiffusionDet)2.1310.1314.273.9142.648.5
DINOv3-T (GOLD)1.4912.3914.273.9155.962.2
DINOv3-S (Baseline: DiffusionDet)1.9111.3211.584.3744.850.9
DINOv3-S (GOLD)1.3414.2811.584.3757.663.9
DINOv3-B (Baseline: DiffusionDet)1.7312.718.715.3646.352.4
DINOv3-B (GOLD)1.1215.538.715.3658.765.1
DINOv3-L (Baseline: DiffusionDet)1.5414.186.606.4247.453.1
DINOv3-L (GOLD)0.9517.626.606.4259.565.6

D. DiffusionDet baseline configuration.

The row labeled “DiffusionDet (ICCV’23)” uses the original Swin-B backbone configuration, with the image resolution, proposal settings, and training schedule same to our experiments. We provide a detailed analysis of the performance of the DiffusionDet baseline equipped with the same backbone as our method in Tables 9WTK 2 and 9WTK 3.

E. Sources of comparison results.

To avoid ambiguity and potential concerns from the research community, and following standard practice in DGOD, the results for prior methods in Tables 1–7 are taken directly from their respective publications.

AuthorsOpenReview ↗

Response to Reviewer 9WTK - Question 2-5

9WTK - Question 2

We confirm that all reported experiments used the proposed method described in the manuscript. The submitted SHEP pipeline implements Section 4.3 and uses generic appearance perturbations without explicitly synthesizing fog or other specific weather conditions.

For PAD, the effective stage weights are [0.05, 0.05, 0.10, 0.20, 0.25, 0.40], as specified in configs/gold_diverse_weather.yaml (line 89). This YAML configuration overrides the defaults in diffusiondet/config.py through cfg.merge_from_file in train_net.py (line 97). This is the same as Table 14 that we provided in Appendix B.

We take transparency and reproducibility seriously and provided the implementation to enable direct verification. We welcome reviewers to run and evaluate the submitted code and inspect its correspondence with the method.

9WTK - Question 3

We report the results from the best-performing checkpoint during training rather than the final checkpoint at 450k iterations. The 450k iteration limit is inherited from the original DiffusionDet codebase and represents the maximum training iterations. In practice, our model typically reaches its peak performance between 100k and 150k iterations. As shown in Table 9WTK 1, the peak iterations across five independent runs are highly consistent and concentrated within a narrow range, indicating stable convergence behavior.

The hyperparameters were selected through experiments on labeled source-domain data. We have provided additional explanations (highlight in blue) in Appendix B.

9WTK - Question 4

As stated in lines 315–316, we use the Foggy Cityscapes split with attenuation coefficient $$\beta=0.02=0.02. Lines 316–317 specify mAP at an IoU threshold of 0.5 (COCO-style AP50, computed through COCOEvaluator in train_net.py lines 26–30). All these settings follow standard practices in DGOD field. Additionally, the default number of DDIM sampling steps is 1.

9WTK - Question 5

A. Additional relevant comparisons.

Given the timeliness of conference research, we prioritize recent methods, including seven methods from 2026 in Table 1, which sufficiently demonstrate the superiority of our GOLD.
Thank you for identifying additional relevant work. We have expanded Table 1, and the additions highlighted in blue.

Table 9WTK 4. New DG results (%) on Diverse Weather Datasets.

MethodVenueDFDRNRNSAvg.
OA-DGAAAI'2438.033.916.838.331.8
SDG-DiffDetICCV'2543.240.944.726.638.9
GOLD (Ours)—55.662.348.860.556.8

B. Per-class results for Bridge and LTFE.

The Bridge and LTFE papers did not provide the per-class results needed for Tables 2–5. We contacted the authors to obtain these results and have incorporated the author-provided values into the updated tables 2-5 (hilghlight in blue).

AuthorsOpenReview ↗

Follow-up on Our Response to Reviewer 9WTK

Dear Reviewer 9WTK,

Thank you again for your careful review. We wanted to follow up on our response.

To address your questions about variability and the contribution of SCAdapter, we reported five runs on Diverse Weather, including the variation in mAP and the iteration at which each run peaked (Table 9WTK 1). We also compared SCAdapter with a linear adapter using the same frozen DINOv3 backbone and with SHEP and PAD disabled. SCAdapter improved average mAP from 46.0 to 48.8 (Table 9WTK 2). We added comparisons on BDD100K and Foggy Cityscapes using matched DINOv3 backbones (Table 9WTK 3), clarified how checkpoints and hyperparameters were selected, and corrected the augmentation description to match the code.

When you have a chance, we would appreciate hearing whether these results and clarifications address your concerns. If any point remains open, we would be glad to discuss it further. Thank you for your time and consideration.

GOLD authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.