Submitted:
17 November 2026
Posted:
18 November 2026
You are already at the latest version
Abstract
Data augmentation is a widely used strategy for single-domain generalization in object detection (S-DGOD). At the same time, it remains unclear whether methods developed primarily for street-view imagery transfer effectively to remote sensing. This study systematically evaluates the augmentation components of nine S-DGOD methods on street-view and remote sensing datasets under natural domain shifts and seven applicable methods under synthetic common-corruption shifts. Natural shifts are assessed using the Diverse-Weather Dataset (DWD) and Real-World Distribution Shifts across Climate Zones (RWDS-CZ). Corruption shifts are evaluated using Cityscapes-C and DOTA-C with matched corruption types and severities. Target-domain performance is measured using mean average precision at an intersection-over-union threshold of 0.5, averaged across target conditions. We use closed-gap metric, because it normalizes augmentation gains relative to source-only and target-trained reference detectors. Under natural shifts, all nine methods showed smaller normalized gains on RWDS-CZ than on DWD. Four of these methods underperformed the source-only baseline on RWDS-CZ. Under common corruptions, five of seven methods achieved higher closed-gap values on DOTA-C than on Cityscapes-C. These findings indicate that augmentation effectiveness depends on both imagery types and shift characteristics, highlighting the importance of evaluating S-DGOD methods on realistic remote sensing domain shifts.
Keywords:
single-domain generalization
; object detection
; data augmentation
; remote sensing
; domain shift
; common corruptions
1. Introduction
Deep neural networks have driven substantial advances in object detection [1,2,3,4,5]. However, object detector performance may deteriorate when the test-data distribution differs from that encountered during training. This phenomenon is known as domain shift [6,7,8]. Several learning paradigms seek to address this challenge. Unsupervised domain adaptation (UDA) adapts a model trained on source domain using unlabeled data from known target domain [9,10,11]. Although effective, UDA may be impractical when representative target-domain data are difficult or costly to obtain [12]. Domain generalization (DG), by contrast, aims to achieve robust performance on unseen target domains by training on multiple labeled source domains. However, acquiring annotated data from several domains may also be infeasible [12,13]. Single-domain generalization (S-DG) addresses a more restrictive setting in which a model must generalize from a single source domain without access to target-domain data during training [14,15,16]. Despite its challenges, S-DG is particularly relevant when only one labeled dataset is available [12,13].
Although S-DG has been studied extensively in image classification, its application to object detection, termed single-domain generalization for object detection (S-DGOD), has emerged as a distinct research area more recently [17,18,19,20,21,22]. Data augmentation is a common strategy for S-DG: generating diverse variants of source-domain images exposes a model to variations that may arise in unseen domains [16,23,24,25]. It is also central to many S-DGOD approaches [20,22]. However, augmentations developed for image classification do not necessarily transfer directly to object detection. They may distort object appearance, impair recognition, or compromise the alignment between objects and their annotated bounding boxes [20]. These challenges have motivated augmentation strategies specifically designed for object detection.
Robust generalization is especially important in remote sensing, where deep learning supports applications such as disaster damage assessment [26], maritime surveillance [27], critical infrastructure monitoring [28], and environmental enforcement [29]. In these applications, object detectors may be deployed in geographic regions or environmental conditions that differ substantially from those represented in the training data. For example, object detectors deployed for disaster response may need to operate across regions with substantially different climatic conditions.
Nevertheless, most S-DGOD augmentation methods are evaluated on street-view benchmarks [17,18,19,20,21,22,30,31,32]. Remote sensing imagery differs from street-view imagery in several respects. Aerial images are typically captured from an overhead perspective and can contain large scenes with objects exhibiting substantial variations in scale, orientation, shape, and spatial density [33]. These characteristics may influence how augmentation affects object appearance, spatial context, and detection performance. Yet, to the best of our knowledge, no systematic comparison has established how the effectiveness of existing S-DGOD augmentation methods differs between street-view and remote sensing imagery.
We address this gap by evaluating S-DGOD augmentation methods on both types of imagery under two forms of distribution shift: naturally occurring domain shifts and synthetic shifts induced by common image corruptions. To facilitate comparisons across datasets, we summarize target-domain performance using mean performance under corruption (mPC), originally introduced for corruption benchmarks by Michaelis et al. [34]. Following the averaging approach previously applied to natural weather domains in [20], we compute the mean mAP@0.5 across target subsets. We then use the closed-gap (CG) metric. CG expresses the performance improvement due to augmentation as a fraction of difference between a source-only detector and reference detector trained on labeled target-domain data. Originally introduced for UDA in 3D object detection [35] and subsequently applied to S-DG in monocular 3D object detection [36], CG provides a normalized measure of target-domain improvement. Here, we apply it to S-DGOD to compare augmentation performance across street-view and remote sensing benchmarks.
Our results reveal a contrast between the two types of domain shift. Under natural shifts, existing S-DGOD augmentation methods are generally less effective on remote sensing imagery than on street-view imagery; some even reduce target-domain performance relative to a source-only baseline. Under common-corruption shifts, however, their effectiveness on remote sensing imagery is generally comparable to or greater than achieved on street-view imagery. One possible explanation is that common-corruption benchmarks are widely used to develop and evaluate S-DGOD methods, potentially favoring techniques that address these shifts across imagery types.
The main contributions of this study are as follows:
- We systematically evaluate S-DGOD augmentation methods on street-view and remote sensing datasets under both natural and corruption-induced domain shifts, using a normalized metric to compare target-domain performances.
- We identify contrasting patterns across the two settings: existing methods yield smaller, and sometimes negative, gains under natural domain shifts in remote sensing imagery. At the same time, their gains under common corruptions are generally comparable to or greater than those obtained on street-view imagery.
2. Related Work
We review representative S-DGOD approaches, focusing on data augmentation, and then discuss related work on domain generalization for object detection in remote sensing imagery.
2.1. S-DGOD
Wu and Deng [17] formalized the S-DGOD setting and introduced a cyclic-disentangled self-distillation framework for learning domain-invariant representations without domain labels. They also introduced the Diverse-Weather Dataset (DWD), which has become a widely used S-DGOD benchmark. Subsequent research has explored several complementary directions. Vidit et al. [18] employed a pretrained vision–language model to generate semantic domain variations through text prompts, combining feature-level semantic augmentation with a text-based classification loss. G-NAS [19] used differentiable neural architecture search and introduced a G-loss to discourage overfitting to easily learned features.
Other methods explicitly seek to preserve object information. OA-DG [20] combines multilevel, semantics-aware transformations with a contrastive objective to mitigate the object-level degradation that can result from classification-oriented augmentation. Liu et al. [21] represented scene-level and object-attribute confounders using a structural causal model and combined Global–Local Transformation (GLT) with Causal Prototype Learning. Danish et al. [22] showed that a carefully selected set of simple augmentations can substantially improve a baseline object detector and complemented this approach with cross-view alignment for classification and localization.
More recent approaches include PhysAug [30], which draws on atmospheric optics to construct frequency-based perturbations approximating real-world imaging effects, and SDG-DiffDet [31], which integrates feature augmentation and target-to-source distribution transfer within a diffusion-based framework. Because data augmentation is central to our experiments, we examine the relevant techniques in greater detail in Section 2.2.
2.2. S-DGOD Data Augmentation Methods
Data augmentation is widely used in S-DGOD because it increases training-image diversity without requiring labeled data from additional domains. Methods differ in the transformations they apply and the mechanisms they use to preserve object information.
Several approaches operate primarily in the image domain. OA-Mix [20] introduces localized domain variations through multilevel transformations and object-aware mixing, with the aim of preserving object appearance and annotation validity. Div [22] applies one transformation per image, selected from a curated collection of existing augmentations. Despite its simplicity, this strategy produced substantial improvements over earlier S-DGOD methods. Color Perturbation (CP) [37] generates color variations by randomly permuting RGB channels.
A second group of methods modifies image frequency representations. TBSA [32] mixes an image’s frequency spectrum with that of a patch sampled from the same image. GLT [21] combines global frequency-domain transformations with spatial-domain transformations of individual objects, using masks generated by the Segment Anything Model (SAM). PhysAug [30] employs physics-inspired frequency-domain perturbations to approximate nonideal imaging conditions. F-LGAM [38] applies Fourier-based augmentation to both local object regions and global image components to simulate varied weather conditions. SCG [39] uses a band-pass mask on discrete cosine transform (DCT) coefficients to separate intermediate frequencies, which are assumed to contain more causal information, from very low and very high frequencies, which are assumed to capture more domain-specific information. SCG then randomly perturbs these very low and very high frequencies to increase domain diversity.
SDAPM+DHDPG [40] follows a different approach, using a two-component augmentation framework based on image statistics to preserve foreground details.
To the best of our knowledge, none of these methods was evaluated on remote sensing imagery in its original publication. Moreover, the remote sensing studies discussed in Section 2.3 do not systematically compare this set of augmentations. GOOD [41] does not evaluate them, whereas AdvSP [42] evaluates SCG on DroneWDD [42,43] but excludes the other methods. Consequently, the transferability of established S-DGOD augmentations to remote sensing remains insufficiently understood.
2.3. S-DGOD in Remote Sensing
Several studies have addressed S-DGOD in remote sensing through feature-level adaptation or spectral representations. GOOD [41] uses CLIP [44] to generate diverse feature-level styles from a single source domain and introduces two loss terms to stabilize object detection across these variations. Thus, GOOD augments intermediate features rather than directly modifying image pixels. Wang et al. [45] introduced learnable filters that decompose an image’s amplitude spectrum into domain-invariant and domain-specific components. The object detector is trained using the invariant component, while an instance-level contrastive objective guides filter learning. AdvSP [42], by contrast, generates adversarial perturbations of both amplitude and phase spectra and trains the object detector on source and perturbed images through min–max optimization. Because these perturbations are learned jointly with the detector, AdvSP is not a standalone method with a fixed augmentation procedure.
Other methods emphasize feature alignment and adaptation. SA-DETR [46] maintains a memory of feature statistics from source and augmented images and uses a domain-style adapter at inference time to shift representations of unseen images toward the stored styles. Its object-aware contrastive component uses class-specific queries and bounding-box masks to align representations of objects from the same class across domains. Bridge [47] introduces a causal module for frozen vision foundation models. It learns low-rank bases for a front-door adjustment designed to reduce the influence of confounding factors, such as illumination and style, while removing redundant features. Related studies also consider synthetic aperture radar (SAR) imagery [48,49,50], which falls outside our focus on visible-spectrum optical imagery.
Together, these studies demonstrate growing interest in remote sensing S-DGOD. However, they primarily emphasize feature-level style variation, spectral disentanglement, causal adjustment, or jointly learned perturbations. Therefore, they do not establish whether broadly applicable image-level augmentation methods retain their effectiveness when transferred from street-view to remote sensing imagery. Our study investigates this unresolved empirical question.
3. Materials and Methods
3.1. Framework for Comparing Out-of-Domain Performance Across Datasets
For each dataset, let denote the source (in-domain) data and the target (out-of-domain) evaluation data. We assess each augmentation method using a source-only baseline, a target-trained reference, and an object detector trained with the augmentation under consideration. Within each dataset, all object detectors share the same architecture and core training protocol; they differ only in their training data and use of the evaluated augmentation:
- : a source-only baseline trained on without the augmentation method under evaluation;
- : an object detector trained on with the augmentation method under evaluation;
- : a target-trained reference object detector trained on labeled target-domain examples without the augmentation method under evaluation.
All three object detectors are evaluated on the same held-out target data. Let , , and denote their respective dataset-level mPC scores on , as defined below. We measure the relative improvement achieved through augmentation using the closed-gap metric introduced by Yang et al. [35]:
The denominator represents the performance gap between the target-trained reference and the source-only baseline, while the numerator represents the gain attributable to augmentation; both are measured on the target data. When , a value of indicates no improvement over the source-only baseline, indicates that the augmented object detector matches the target-trained reference, and indicates that augmentation reduces target-domain performance. Values above one are possible when the augmented object detector outperforms the reference. The metric is undefined if and should be interpreted cautiously when the reference gap is very small or negative.
By normalizing improvements against the corresponding reference gap, CG facilitates comparisons across datasets with different absolute performance levels. For reporting, we express CG values as percentages.
3.2. Experimental Setup
We evaluate nine S-DGOD augmentation methods on four datasets. Two datasets capture natural domain shifts: the Diverse-Weather Dataset (DWD) [17], which contains street-view images acquired under different weather and lighting conditions, and Real-World Distribution Shifts across Climate Zones (RWDS-CZ) [51], which contains remote sensing imagery from different climate zones. The other two datasets capture synthetic shifts induced by common image corruptions: Cityscapes-C [34], based on street-view imagery, and DOTA-C [52], based on overhead remote sensing imagery. The evaluated augmentation methods are OA-Mix [20], SDAPM+DHDPG [40], GLT [21], SCG [39], TBSA [32], PhysAug [30], CP [37], Div [22], and F-LGAM [38]. All experiments used Faster R-CNN detectors, with dataset-specific architectures and training configurations detailed in Appendix A.
For each dataset, we train a source-only object detector , a target-trained reference , and a separate augmented object detector for each applicable method. Within each dataset, we hold the architecture and core training protocol constant and evaluate all object detectors on identical held-out target subsets. We first measure object detection performance within each subset using mean average precision at an intersection-over-union (IoU) threshold of 0.5 (mAP@0.5), denoted by for model M evaluated on subset D. We then report mean performance under corruption (mPC), following the averaging principle introduced by Michaelis et al. [34]. Specifically, for benchmark B with a set of target evaluation conditions, we define
Each target condition contributes equally, regardless of the number of images in its subset. For corruption benchmarks, a condition corresponds to a corruption type and severity level; for natural-shift benchmarks, we follow the averaging approach used in [20] and compute the unweighted mean of mAP@0.5 across distinct target domains. Thus, mPC here always refers to averaged mAP@0.5 rather than to an average over multiple IoU thresholds. We substitute the resulting dataset-level mPC scores into CG, rather than averaging condition-wise CG values.
We match object detector architectures and training protocols as closely as possible within each comparison pair—DWD versus RWDS-CZ and Cityscapes-C versus DOTA-C—while allowing adjustments required by dataset-specific characteristics. This design aims to reduce confounding differences between the street-view and remote sensing experiments.
3.2.1. DWD Versus RWDS-CZ: Natural Domain Shift
DWD comprises five subsets: Daytime-Sunny, Night-Sunny, Dusk-Rainy, Night-Rainy, and Daytime-Foggy. We designate Daytime-Sunny as the source domain and the remaining four subsets as target domains. The source-only and augmented object detectors are trained on the Daytime-Sunny training split. The target-trained reference is trained on the combined training splits of Night-Sunny, Dusk-Rainy, Night-Rainy, and Daytime-Foggy, without the augmentation method under evaluation. A separate is trained for each augmentation method.
RWDS-CZ comprises three subsets: CZ-A, CZ-B, and CZ-C. We use CZ-A as the source domain and CZ-B and CZ-C as target domains. The source-only and augmented object detectors are trained on the CZ-A training split, whereas the reference detector is trained on the combined training splits of CZ-B and CZ-C. As with DWD, we train a separate augmented object detector for each method. Further details on detector configurations, data preprocessing, training schedules, and dataset splits for DWD and RWDS-CZ are provided in Appendices Appendix A.1 and Appendix A.2, respectively.
For DWD, let consist of the four target subsets Night-Sunny, Dusk-Rainy, Night-Rainy, and Daytime-Foggy. For RWDS-CZ, let consist of CZ-B and CZ-C. Writing for the mAP@0.5 on target subset t, their mPC scores are
For both datasets, all models are evaluated on the same held-out images from the respective target subsets. Object detector architectures and training protocols are kept as similar as possible across DWD and RWDS-CZ, subject to dataset-specific requirements.
3.2.2. Cityscapes-C Versus DOTA-C: Synthetic Common-Corruption Shift
Cityscapes-C and DOTA-C are derived from Cityscapes [53] and DOTA-v1.0 [33], respectively, by applying synthetic image corruptions. The original Cityscapes-C benchmark evaluates 15 corruption types at five severity levels [34], whereas the full DOTA-C benchmark provides 19 types at five severity levels, as well as a separate cloud-corruption benchmark [52]. For the matched Cityscapes-C versus DOTA-C comparison considered here, we apply all 19 corruption types from DOTA-C to both datasets, each at severity levels . Thus, our Cityscapes-C evaluation extends the original 15-type corruption set to the 19-type DOTA-C set. The separate cloud-corruption benchmark is not included in this comparison. In each benchmark, the original clean images constitute the source domain, and their corrupted counterparts define the target evaluation conditions. The source-only object detector is trained on clean images, while each augmented object detector is trained on clean images with its respective S-DGOD augmentation method. The reference detector is trained on source training images modified using corruptions drawn from the evaluation corruption pool, thereby providing a corruption-aware reference under the same annotation scheme.
Let denote these 19 corruption types and let denote the held-out evaluation images from benchmark B with corruption c at severity s. With , the mPC for each of the two benchmarks is
where B is Cityscapes-C or DOTA-C, respectively. In both cases, mPC averages 95 corruption–severity conditions, with equal weighting across corruption types and severity levels.
We evaluate all selected augmentation methods on both benchmarks except SDAPM+DHDPG and Div. We exclude these two methods because their augmentation sets overlap with the corruptions defining the target evaluation conditions. All models are tested on identical held-out corrupted images. As in the natural-shift experiments, we align object detector architectures and core training protocols across Cityscapes-C and DOTA-C while accommodating dataset-specific requirements. Additional implementation details, including image preprocessing, training configurations, and evaluation procedures for Cityscapes-C and DOTA-C, are provided in Appendices Appendix A.3 and Appendix A.4, respectively.
4. Results
4.1. DWD Versus RWDS-CZ: Natural Domain Shift
Table 1 reports the target-domain mPC scores (mean mAP@0.5 across target subsets) of the source-only and target-trained reference detectors on DWD and RWDS-CZ, as defined in Formulas (3)-(). On DWD, achieved 26.2%, whereas achieved 41.1%, yielding a reference gap of 14.9 percentage points. On RWDS-CZ, the corresponding scores were 8.8% and 16.4%, yielding a gap of 7.6 percentage points. Because both reference gaps are positive, positive CG values indicate improvements over the respective source-only baselines.
Table 2 reports the CG values, defined by Formula (1), for all nine augmentation methods. Every method achieved a higher CG on DWD than on RWDS-CZ. To assess the cross-dataset decreases in CG, the last column shows the difference between DWD and RWDS-CZ (DWD minus RWDS-CZ). The differences ranged from 0.8 percentage points for GLT to 22.9 percentage points for Div. On DWD, all methods outperformed the source-only baseline: CG values ranged from 4.2% for CP to 34.2% for Div, with PhysAug ranking second at 29.7%. Results on RWDS-CZ were more varied. PhysAug achieved the highest CG (12.7%), followed by Div (11.3%) and GLT (9.3%). Four methods yielded negative values: SDAPM+DHDPG (%), SCG (%), TBSA (%), and CP (%). Thus, these four methods reduced mean target-domain detection performance relative to the source-only baseline on RWDS-CZ, despite improving performance on DWD. Detailed source- and target-domain mAP@0.5 results for individual weather and lighting conditions in DWD and climate zones in RWDS-CZ are provided in Appendix B.1.
GLT, F-LGAM, and OA-Mix showed the smallest cross-dataset decreases in CG (0.8, 2.6, and 4.2 percentage points, respectively). These methods distinguish object-level transformations from broader image- or background-level transformations. Their comparatively small decreases are consistent with the possibility that preserving object appearance during augmentation is particularly important in remote sensing imagery, where objects may occupy relatively few pixels. However, this observational pattern does not establish which augmentation components account for the differences.
4.2. Cityscapes-C Versus DOTA-C: Synthetic Common-Corruption Shift
Table 3 shows the mPC scores (mean mAP@0.5 over the evaluated corruption–severity conditions), defined by Formula (5). On Cityscapes-C, the source-only and corruption-aware reference detectors achieved 24.8% and 60.7%, respectively, yielding a reference gap of 35.9 percentage points. On DOTA-C, the corresponding scores were 37.6% and 54.3%, yielding a gap of 16.7 percentage points. Both reference gaps were positive, as in the natural-shift experiments.
Table 4 presents the CG values for the seven applicable augmentation methods. In contrast to the natural-shift experiments, five of the seven methods achieved higher CG values on the remote sensing benchmark DOTA-C than on the street-view benchmark Cityscapes-C. The largest advantages on DOTA-C were observed for PhysAug (7.6 percentage points), SCG (7.1 percentage points), and GLT (4.6 percentage points). TBSA also performed better on DOTA-C, while OA-Mix yielded nearly identical values on the two datasets (12.1% and 12.2%). F-LGAM was more effective on Cityscapes-C (18.4%) than on DOTA-C (15.0%). CP was the only method with negative CG on either benchmark, yielding % on Cityscapes-C and % on DOTA-C. Unlike in the natural-shift experiments, no method improved performance on the street-view benchmark while degrading it on the remote sensing benchmark.
On DOTA-C, PhysAug achieved the highest CG (33.4%), followed by GLT (31.5%). These two methods were also the strongest performers on Cityscapes-C, although their ranking was reversed: GLT achieved 26.9%, compared with 25.8% for PhysAug. Overall, the results suggest that, for the evaluated common corruptions, the tested augmentations were generally at least as effective at closing the reference gap in remote sensing imagery as in street-view imagery. Detailed results for individual corruption types, including mAP@0.5 averaged across the five severity levels and clean-image detection performance, are reported in Appendix B.2.
4.3. Comparison of Natural and Corruption-Induced Domain Shifts
The two experimental settings revealed contrasting cross-dataset patterns. Under natural shifts, all nine augmentations achieved lower CG values on RWDS-CZ than on DWD, and four reduced performance below the RWDS-CZ source-only baseline. Under common-corruption shifts, five of the seven augmentations achieved higher CG values on DOTA-C than on Cityscapes-C, and CP was the only method with a negative CG on either benchmark. Furthermore, GLT, F-LGAM, and OA-Mix, which exhibited comparatively small cross-dataset decreases under natural shifts, did not form a similarly distinctive group under common-corruption shifts.
The relationship between street-view and remote sensing performance also differed by shift type. Across augmentation methods, the Pearson correlation between CG values on DWD and RWDS-CZ was , compared with for Cityscapes-C and DOTA-C. Thus, augmentation performance on street-view imagery was more strongly associated with performance on remote sensing imagery under synthetic common corruptions than under natural shifts. These correlations are descriptive and should be interpreted in light of the limited number of evaluated methods.
One possible explanation is that common-corruption benchmarks are widely used in the development and evaluation of S-DGOD augmentation methods, whereas natural geographic and environmental shifts in remote sensing are less directly represented in such evaluations. Our experiments do not establish a causal explanation for the observed differences. Nevertheless, they demonstrate that conclusions about augmentation transferability depend strongly on the type of domain shift: favorable results on common-corruption benchmarks do not necessarily translate into improved performance under natural shifts in remote sensing imagery.
5. Discussion
The results show that the transferability of S-DGOD data augmentation from street-view to remote sensing imagery depends strongly on the type of domain shift. Under natural shifts, all nine methods closed a smaller fraction of the source-to-reference performance gap on RWDS-CZ than on DWD, and four methods decreased performance relative to the RWDS-CZ source-only baseline. Under synthetic common corruptions, by contrast, five of the seven applicable methods achieved higher CG values on DOTA-C than on Cityscapes-C. These contrasting outcomes indicate that performance on a corruption benchmark alone is not sufficient evidence that an augmentation will improve generalization to naturally occurring shifts in remote sensing imagery. They also qualify the conclusions that can be drawn from the strong performance of augmentation-based approaches on established street-view benchmarks.
Differences among methods provide additional, although not conclusive, insight. GLT, F-LGAM, and OA-Mix exhibited the smallest cross-dataset decreases in CG under natural shifts. Their use of local or object-aware transformations alongside broader image-level changes is consistent with the possibility that controlling perturbations around objects helps preserve detection-relevant information. Nevertheless, this pattern does not persist as a distinctive ranking under common corruptions.
The CG metric makes cross-benchmark comparisons more interpretable by expressing augmentation gains relative to a dataset-specific reference gap, but it should not be confused with an absolute improvement in mAP. The reference gaps were 14.9 and 7.6 percentage points for DWD and RWDS-CZ, and 35.9 and 16.7 percentage points for Cityscapes-C and DOTA-C, respectively. For example, GLT closed a larger fraction of the gap on DOTA-C than on Cityscapes-C (31.5% versus 26.9%), even though these values correspond to approximate absolute mPC gains of 5.3 and 9.7 percentage points, respectively.
6. Conclusions
This study compared nine S-DGOD augmentation methods on street-view and remote sensing imagery under natural domain shifts and, for seven applicable methods, synthetic common-corruption shifts. Using mean target-domain mAP@0.5 and the closed-gap metric, we found that all nine methods achieved lower normalized gains on the remote sensing benchmark RWDS-CZ than on the street-view benchmark DWD; four methods even reduced RWDS-CZ performance below the source-only baseline. In contrast, five of seven methods attained higher CG values on DOTA-C than on Cityscapes-C under matched common corruptions.
These findings demonstrate that the apparent transferability of augmentation methods depends on how out-of-domain performance is evaluated. Strong results on street-view data or synthetic corruptions do not necessarily predict improvements under natural geographic and environmental shifts in remote sensing. Future S-DGOD research should therefore prioritize evaluation on realistic target domains and investigate augmentation designs that expand domain diversity without compromising object information.
Author Contributions
Conceptualization, T.C.; methodology, T.C.; software, T.C.; formal analysis, T.C.; investigation, T.C.; writing—original draft preparation, T.C.; writing—review and editing, T.C. and O.I.; supervision, O.I. and M.T. All authors have read and agreed to the published version of the manuscript.
Funding
This research received no external funding.
Institutional Review Board Statement
Not applicable.
Informed Consent Statement
Not applicable.
Data Availability Statement
Conflicts of Interest
The authors declare no conflicts of interest.
Abbreviations
The following abbreviations are used in this manuscript:
| CG | Closed-Gap |
| CLIP | Contrastive Language–Image Pretraining |
| COCO | Common Objects in Context |
| CP | Color Perturbation |
| DCT | Discrete Cosine Transform |
| DG | Domain Generalization |
| DWD | Diverse-Weather Dataset |
| FPN | Feature Pyramid Network |
| GLT | Global–Local Transformation |
| GPU | Graphics Processing Unit |
| IoU | Intersection over Union |
| mAP | Mean Average Precision |
| mPC | Mean Performance under Corruption |
| R-CNN | Region-based Convolutional Neural Network |
| RGB | Red, Green, and Blue |
| RoI | Region of Interest |
| RWDS-CZ | Real-World Distribution Shifts across Climate Zones |
| SAM | Segment Anything Model |
| SAR | Synthetic Aperture Radar |
| S-DG | Single-Domain Generalization |
| S-DGOD | Single-Domain Generalization for Object Detection |
| UDA | Unsupervised Domain Adaptation |
Appendix A. Object Detector Architectures and Training Settings
All experiments used horizontal-bounding-box Faster R-CNN detectors [3]. Within each dataset, the source-only detector , augmented detectors , and target-trained reference shared the same architecture, initialization procedure, optimizer, and training schedule. Each augmented object detector added the corresponding image augmentation to the common training pipeline; auxiliary losses and architectural modules associated with the original methods were not included. Consequently, the reported results characterize the augmentation components under the experimental protocol, rather than complete reproductions of the original methods.
Table A1 summarizes the dataset-specific settings. All models used stochastic gradient descent with momentum and weight decay . The learning rate increased linearly during the first 500 iterations, starting from times its initial value, and was subsequently multiplied by at each listed milestone. Training used one GPU per model. The common augmentation pipeline included resizing and random horizontal flipping with probability . Method-specific augmentations were applied after these operations.
Table A1.
Object detector architectures and training settings. Scheduler epochs refer to epochs over the possibly repeated training dataset. Effective passes equal scheduler epochs multiplied by the dataset repetition factor. Learning-rate milestones are expressed in scheduler epochs. Gradient clipping, where used, limits the global gradient norm.
Table A1.
Object detector architectures and training settings. Scheduler epochs refer to epochs over the possibly repeated training dataset. Effective passes equal scheduler epochs multiplied by the dataset repetition factor. Learning-rate milestones are expressed in scheduler epochs. Gradient clipping, where used, limits the global gradient norm.
| Setting | DWD | RWDS-CZ | Cityscapes-C | DOTA-C |
|---|---|---|---|---|
| Detector | Faster R-CNN | Faster R-CNN | Faster R-CNN | Faster R-CNN |
| Backbone | ResNet-101-DC5 | ResNet-101 | ResNet-50 | ResNet-50 |
| Feature pyramid | None | FPN | FPN | FPN |
| Number of classes | 7 | 16 | 8 | 15 |
| Initialization | ImageNet backbone | COCO detector | COCO detector | COCO detector |
| Training batch size | 2 | 2 | 1 | 2 |
| Initial learning rate | 0.001 | 0.0025 | 0.00125 | 0.000625 |
| Scheduler epochs | 10 | 12 | 8 | 12 |
| Dataset repetition factor | 2 | 3 | 8 | 1 |
| Effective training passes | 20 | 36 | 64 | 12 |
| Learning-rate milestones | 4, 8 | 9, 11 | 7 | 8, 11 |
| Gradient-norm limit | None | 35 | None | 35 |
The FPN detectors used 256-channel pyramid features, a region proposal network, RoIAlign features, and a bounding-box head with two 1024-unit fully connected layers. Their backbones used the PyTorch-style ResNet implementation. The stem and first residual stage were frozen, and batch-normalization running statistics were held fixed. Images were converted to RGB and normalized.
Appendix A.1. DWD
DWD used a Caffe-style ResNet-101 backbone with a dilated final residual stage (DC5), without an FPN. The final backbone feature map had an output stride of 16 and 2048 channels. This feature map supplied both the region proposal network and the RoI feature extractor. The RoI head used RoIAlign followed by two 1024-unit fully connected layers. The backbone was initialized from the ImageNet-pretrained weights; the detection heads were trained for the seven DWD classes. The stem and first residual stage were frozen, and batch-normalization parameters and running statistics were fixed.
Training and evaluation images were resized with preserved aspect ratio to fit within a width of 1000 pixels and a height of 600 pixels. The training dataset was repeated twice during each of the ten scheduler epochs, giving 20 effective passes.
The source-only and augmented detectors used the official Daytime-Sunny training split, and source-domain evaluation used its official test split. Each target domain was deterministically divided into an approximately 80% training portion and a 20% evaluation portion. The four training portions were pooled to train ; the evaluation portions were retained separately for Night-Sunny, Dusk-Rainy, Night-Rainy, and Daytime-Foggy. These evaluation images were excluded from reference training.
Appendix A.2. RWDS-CZ
RWDS-CZ used a ResNet-101-FPN detector initialized from a COCO-pretrained Faster R-CNN checkpoint, with the classification and regression outputs adapted to its 16 classes. The input data consisted of the released tiles. During training, aspect-ratio-preserving random resizing used the scale interval defined by and , where each pair specifies width and height. Evaluation used the fixed scale .
The training dataset was repeated three times per scheduler epoch, giving 36 effective passes over 12 scheduler epochs. Background-only training tiles were retained: no empty-ground-truth filtering was applied. The source-only and augmented object detectors were trained on CZ-A. The target-trained reference used the union of the released CZ-B and CZ-C training splits. Source and target evaluations used the corresponding released test splits.
Appendix A.3. Cityscapes-C
Cityscapes-C used a ResNet-50-FPN Faster R-CNN initialized from a COCO-pretrained object detector and adapted to eight classes. Training used aspect-ratio-preserving random resizing between and ; evaluation used . The training dataset was repeated eight times per scheduler epoch. Thus, eight scheduler epochs corresponded to 64 effective passes, with the learning-rate reduction after seven scheduler epochs.
The source-only and augmented object detectors used clean Cityscapes training images. For , each training-image presentation received one uniformly sampled corruption type and one uniformly sampled severity from the same 19-type, five-severity pool used for evaluation. Evaluation used the 500 Cityscapes validation images and their corrupted counterparts.
Appendix A.4. DOTA-C
DOTA-C used the same ResNet-50-FPN detector architecture and COCO-pretrained initialization as Cityscapes-C, with outputs adapted to the 15 DOTA-v1.0 classes. Original images were divided into tiles with 200-pixel overlap. Oriented annotations were converted to enclosing horizontal bounding boxes. Both training and evaluation used the fixed scale with preserved aspect ratio. Training lasted 12 epochs without dataset repetition.
The source-only and augmented object detectors used clean training tiles. The corruption-aware reference sampled one corruption type and severity for each training-image presentation, as in Cityscapes-C. For corruption evaluation, a fixed subset of 1200 tiles was sampled from the 5297 validation tiles. The source-domain scores reported below were evaluated on these same 1200 tiles, ensuring that clean and corrupted results refer to identical images.
For DWD, Cityscapes-C, and DOTA-C, training excluded images with empty ground-truth annotations or a shortest image side below 32 pixels. Evaluation used deterministic resizing without the training-time augmentations. All reported scores correspond to the final scheduled checkpoint: epoch 10 for DWD, epoch 8 for Cityscapes-C, and epoch 12 for RWDS-CZ and DOTA-C.
Appendix B. Source- and Target-Domain Detection Performance
Table A2–Table A5 report detection performance for the models included in the main comparisons. All values are percentages. Source-domain columns contain mAP@0.5 on held-out source images, rather than training-set performance. Rows labeled by an augmentation name denote trained with that augmentation.
For natural shifts, each target column reports mAP@0.5 on the corresponding held-out domain. For synthetic shifts, each corruption column reports mAP@0.5 averaged over its five severity levels:
The mPC column averages target conditions only; source-domain performance is excluded. For Cityscapes-C and DOTA-C, mPC therefore averages all 95 corruption–severity combinations.
Appendix B.1. Natural Domain Shifts
Table A2.
DWD source- and target-domain mAP@0.5 (%). Daytime-Sunny is the source domain. Target mPC averages the remaining four domains with equal weights.
Table A2.
DWD source- and target-domain mAP@0.5 (%). Daytime-Sunny is the source domain. Target mPC averages the remaining four domains with equal weights.
| Model | Daytime-Sunny | Night-Sunny | Dusk-Rainy | Night-Rainy | Daytime-Foggy | Target mPC |
|---|---|---|---|---|---|---|
| 49.7 | 33.7 | 25.9 | 14.9 | 30.3 | 26.2 | |
| OA-Mix | 49.7 | 35.2 | 28.5 | 16.6 | 31.0 | 27.8 |
| SDAPM+DHDPG | 48.0 | 32.1 | 31.9 | 16.3 | 29.7 | 27.5 |
| GLT | 49.4 | 35.2 | 27.6 | 16.3 | 31.7 | 27.7 |
| SCG | 49.4 | 36.0 | 27.4 | 17.1 | 31.9 | 28.1 |
| TBSA | 49.5 | 35.2 | 28.6 | 14.7 | 32.9 | 27.9 |
| PhysAug | 49.3 | 36.0 | 33.2 | 21.7 | 31.6 | 30.6 |
| CP | 49.4 | 33.8 | 27.3 | 15.4 | 30.8 | 26.8 |
| F-LGAM | 50.1 | 34.7 | 28.8 | 16.1 | 31.1 | 27.7 |
| Div | 48.6 | 36.0 | 34.7 | 22.0 | 32.5 | 31.3 |
| 44.5 | 42.9 | 42.3 | 38.6 | 40.6 | 41.1 |
Table A3.
RWDS-CZ source- and target-domain mAP@0.5 (%). CZ-A is the source domain. Target mPC averages CZ-B and CZ-C with equal weights.
Table A3.
RWDS-CZ source- and target-domain mAP@0.5 (%). CZ-A is the source domain. Target mPC averages CZ-B and CZ-C with equal weights.
| Model | CZ-A | CZ-B | CZ-C | Target mPC |
|---|---|---|---|---|
| 15.7 | 9.9 | 7.8 | 8.8 | |
| OA-Mix | 16.6 | 10.1 | 8.6 | 9.3 |
| SDAPM+DHDPG | 14.5 | 9.5 | 8.1 | 8.8 |
| GLT | 16.5 | 10.5 | 8.6 | 9.6 |
| SCG | 15.0 | 9.5 | 7.6 | 8.6 |
| TBSA | 15.8 | 10.0 | 7.6 | 8.8 |
| PhysAug | 16.1 | 10.6 | 9.0 | 9.8 |
| CP | 15.8 | 9.4 | 7.6 | 8.5 |
| F-LGAM | 16.5 | 10.3 | 8.5 | 9.4 |
| Div | 16.2 | 10.5 | 8.9 | 9.7 |
| 9.4 | 15.9 | 16.8 | 16.4 |
Appendix B.2. Synthetic Common-Corruption Shifts
Table A4 and Table A5 include the seven augmentation methods retained in the main common-corruption comparison. SDAPM+DHDPG and Div are excluded because their augmentation sets overlap with the target corruptions.
The abbreviated corruption names are GN: Gaussian noise; SN: shot noise; IN: impulse noise; SpN: speckle noise; DB: defocus blur; GlB: glass blur; MB: motion blur; ZB: zoom blur; GB: Gaussian blur; Br: brightness; Sp: spatter; Ctr: contrast; El: elastic transformation; Px: pixelation; JPEG: JPEG compression; and Sat: saturation.
Table A4.
Cityscapes-C source- and target-domain mAP@0.5 (%). Clean denotes source-domain performance. Each corruption score averages five severity levels; target mPC averages all 95 conditions. The two panels describe the same models.
Table A4.
Cityscapes-C source- and target-domain mAP@0.5 (%). Clean denotes source-domain performance. Each corruption score averages five severity levels; target mPC averages all 95 conditions. The two panels describe the same models.
| Model | Clean | GN | SN | IN | SpN | DB | GlB | MB | ZB | GB | Snow |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 66.7 | 1.4 | 2.2 | 1.3 | 10.4 | 31.4 | 30.4 | 30.1 | 3.2 | 34.7 | 2.8 | |
| OA-Mix | 66.6 | 7.0 | 9.7 | 7.7 | 24.0 | 31.2 | 29.8 | 29.1 | 3.7 | 34.1 | 2.3 |
| GLT | 67.8 | 4.5 | 6.5 | 5.4 | 17.6 | 50.9 | 25.8 | 37.0 | 6.7 | 53.3 | 3.0 |
| SCG | 65.7 | 0.7 | 1.2 | 0.2 | 8.3 | 23.8 | 21.0 | 27.3 | 2.8 | 28.0 | 2.6 |
| TBSA | 68.1 | 0.2 | 0.4 | 0.1 | 5.4 | 39.3 | 14.2 | 32.0 | 7.6 | 43.0 | 2.0 |
| PhysAug | 67.0 | 16.6 | 20.4 | 6.4 | 35.0 | 35.6 | 34.4 | 42.7 | 6.7 | 39.3 | 4.0 |
| CP | 64.9 | 1.3 | 2.0 | 0.9 | 10.1 | 24.6 | 23.5 | 25.9 | 2.6 | 28.1 | 1.5 |
| F-LGAM | 68.0 | 18.0 | 22.1 | 14.2 | 37.4 | 28.4 | 28.5 | 29.4 | 3.3 | 32.5 | 2.6 |
| 69.0 | 53.1 | 56.2 | 54.5 | 61.8 | 58.9 | 59.9 | 61.5 | 46.5 | 60.3 | 58.5 |
| Model | Frost | Fog | Br | Sp | Ctr | El | Px | JPEG | Sat | Target mPC |
|---|---|---|---|---|---|---|---|---|---|---|
| 14.0 | 44.4 | 43.3 | 30.7 | 36.5 | 64.1 | 28.7 | 23.9 | 38.3 | 24.8 | |
| OA-Mix | 20.3 | 53.5 | 53.8 | 34.2 | 42.1 | 63.5 | 27.8 | 29.4 | 51.2 | 29.2 |
| GLT | 30.3 | 58.5 | 58.0 | 44.2 | 54.8 | 64.6 | 35.2 | 36.5 | 62.3 | 34.5 |
| SCG | 29.7 | 59.3 | 63.5 | 28.1 | 50.7 | 62.4 | 29.2 | 29.6 | 61.1 | 27.9 |
| TBSA | 42.3 | 63.7 | 59.7 | 24.5 | 57.4 | 64.8 | 36.2 | 24.0 | 55.0 | 30.1 |
| PhysAug | 23.5 | 49.1 | 57.9 | 33.9 | 39.7 | 65.2 | 48.3 | 32.5 | 56.4 | 34.1 |
| CP | 14.6 | 44.9 | 43.3 | 28.5 | 37.3 | 61.5 | 24.0 | 22.9 | 41.8 | 23.1 |
| F-LGAM | 25.3 | 49.2 | 55.4 | 30.2 | 36.4 | 64.0 | 34.1 | 33.4 | 52.8 | 31.4 |
| 61.3 | 67.1 | 67.2 | 63.7 | 65.5 | 67.2 | 64.7 | 59.8 | 66.4 | 60.7 |
Table A5.
DOTA-C source- and target-domain mAP@0.5 (%). Clean and corrupted evaluations use the same fixed subset of 1200 validation tiles. Each corruption score averages five severity levels; target mPC averages all 95 conditions. The two panels describe the same models.
Table A5.
DOTA-C source- and target-domain mAP@0.5 (%). Clean and corrupted evaluations use the same fixed subset of 1200 validation tiles. Each corruption score averages five severity levels; target mPC averages all 95 conditions. The two panels describe the same models.
| Model | Clean | GN | SN | IN | SpN | DB | GlB | MB | ZB | GB | Snow |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 65.9 | 26.4 | 26.9 | 22.8 | 34.8 | 38.7 | 37.8 | 37.2 | 10.8 | 40.7 | 22.7 | |
| OA-Mix | 63.3 | 26.0 | 27.5 | 25.6 | 36.6 | 41.1 | 40.5 | 38.4 | 12.2 | 42.4 | 24.3 |
| GLT | 65.5 | 31.2 | 31.6 | 29.8 | 39.8 | 49.8 | 40.2 | 41.5 | 13.7 | 52.8 | 25.7 |
| SCG | 64.2 | 28.1 | 29.1 | 25.9 | 37.6 | 38.9 | 36.8 | 36.8 | 11.0 | 41.7 | 27.4 |
| TBSA | 65.6 | 25.3 | 26.0 | 21.4 | 33.2 | 42.2 | 35.6 | 41.6 | 17.2 | 45.8 | 28.3 |
| PhysAug | 65.7 | 33.1 | 33.9 | 28.9 | 42.6 | 42.5 | 46.1 | 47.4 | 15.6 | 45.4 | 30.0 |
| CP | 64.9 | 25.6 | 26.3 | 22.5 | 34.6 | 36.6 | 34.1 | 35.3 | 10.8 | 39.0 | 19.9 |
| F-LGAM | 65.3 | 32.3 | 33.4 | 31.2 | 39.3 | 39.6 | 40.1 | 38.5 | 12.0 | 41.8 | 23.3 |
| 62.2 | 50.2 | 51.6 | 49.9 | 55.3 | 53.7 | 55.5 | 55.7 | 36.5 | 55.6 | 49.6 |
| Model | Frost | Fog | Br | Sp | Ctr | El | Px | JPEG | Sat | Target mPC |
|---|---|---|---|---|---|---|---|---|---|---|
| 33.8 | 45.6 | 58.8 | 39.9 | 36.7 | 58.6 | 45.9 | 42.9 | 53.4 | 37.6 | |
| OA-Mix | 34.6 | 47.9 | 59.6 | 46.7 | 41.9 | 57.5 | 46.8 | 47.2 | 56.3 | 39.6 |
| GLT | 38.6 | 53.9 | 59.8 | 42.2 | 47.4 | 59.2 | 50.2 | 47.6 | 59.7 | 42.9 |
| SCG | 42.0 | 54.5 | 61.9 | 39.2 | 47.7 | 57.9 | 45.6 | 43.2 | 58.7 | 40.2 |
| TBSA | 42.9 | 53.3 | 61.0 | 42.3 | 46.9 | 60.4 | 45.0 | 43.1 | 57.1 | 40.5 |
| PhysAug | 41.0 | 49.9 | 63.0 | 42.4 | 40.5 | 60.4 | 54.7 | 47.0 | 55.9 | 43.2 |
| CP | 31.8 | 45.7 | 57.9 | 38.9 | 38.4 | 57.3 | 43.4 | 41.6 | 55.4 | 36.6 |
| F-LGAM | 36.8 | 46.7 | 61.4 | 39.8 | 38.2 | 58.8 | 48.4 | 45.4 | 55.2 | 40.1 |
| 50.6 | 59.4 | 61.4 | 55.3 | 57.4 | 60.3 | 60.2 | 54.5 | 58.8 | 54.3 |
References
- Zou, Z.; Chen, K.; Shi, Z.; Guo, Y.; Ye, J. Object detection in 20 years: A survey. Proc. IEEE 2023, 111, 257–276. [Google Scholar] [CrossRef]
- Zhao, Z.Q.; Zheng, P.; Xu, S.T.; Wu, X. Object detection with deep learning: A review. IEEE Trans. Neural Netw. Learn. Syst. 2019, 30, 3212–3232. [Google Scholar] [CrossRef] [PubMed]
- Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards real-time object detection with region proposal networks. Adv. Neural Inf. Process. Syst. 2015, 28, 91–99. [Google Scholar]
- Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016; pp. 779–788. [Google Scholar]
- Carion, N.; Massa, F.; Synnaeve, G.; Usunier, N.; Kirillov, A.; Zagoruyko, S. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision (ECCV), 2020; Springer; pp. 213–229. [Google Scholar]
- Recht, B.; Roelofs, R.; Schmidt, L.; Shankar, V. Do ImageNet classifiers generalize to ImageNet? In Proceedings of the International Conference on Machine Learning (ICML). PMLR, 2019; pp. 5389–5400. [Google Scholar]
- Koh, P.W.; Sagawa, S.; Marklund, H.; Xie, S.M.; Zhang, M.; Balsubramani, A.; Hu, W.; Yasunaga, M.; Phillips, R.L.; Gao, I.; et al. WILDS: A benchmark of in-the-wild distribution shifts. In Proceedings of the International Conference on Machine Learning (ICML). PMLR, 2021; pp. 5637–5664. [Google Scholar]
- Hendrycks, D.; Dietterich, T. Benchmarking neural network robustness to common corruptions and perturbations. In Proceedings of the International Conference on Learning Representations (ICLR), 2019. [Google Scholar]
- Ganin, Y.; Lempitsky, V. Unsupervised domain adaptation by backpropagation. In Proceedings of the International Conference on Machine Learning (ICML). PMLR, 2015; pp. 1180–1189. [Google Scholar]
- Long, M.; Cao, Y.; Wang, J.; Jordan, M. Learning transferable features with deep adaptation networks. In Proceedings of the International Conference on Machine Learning (ICML), 2015; PMLR; pp. 97–105. [Google Scholar]
- Tzeng, E.; Hoffman, J.; Saenko, K.; Darrell, T. Adversarial discriminative domain adaptation. In Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2017; pp. 2962–2971. [Google Scholar]
- Zhou, K.; Liu, Z.; Qiao, Y.; Xiang, T.; Loy, C.C. Domain generalization: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 2023, 45, 4396–4415. [Google Scholar] [CrossRef] [PubMed]
- Wang, J.; Lan, C.; Liu, C.; Ouyang, Y.; Qin, T.; Lu, W.; Chen, Y.; Zeng, W.; Yu, P.S. Generalizing to unseen domains: A survey on domain generalization. IEEE Trans. Knowl. Data Eng. 2023, 35, 8052–8072. [Google Scholar] [CrossRef]
- Qiao, F.; Zhao, L.; Peng, X. Learning to learn single domain generalization. In Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2020; pp. 12553–12562. [Google Scholar]
- Fan, X.; Wang, Q.; Ke, J.; Yang, F.; Gong, B.; Zhou, M. Adversarially adaptive normalization for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021; pp. 8208–8217. [Google Scholar]
- Wang, Z.; Luo, Y.; Qiu, R.; Huang, Z.; Baktashmotlagh, M. Learning to diversify for single domain generalization. In Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV); IEEE, 2021; pp. 814–823. [Google Scholar]
- Wu, A.; Deng, C. Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation. In Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2022; pp. 837–846. [Google Scholar]
- Vidit, V.; Engilberge, M.; Salzmann, M. CLIP the gap: A single domain generalization approach for object detection. In Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2023; pp. 3219–3229. [Google Scholar]
- Wu, F.; Gao, J.; Hong, L.; Wang, X.; Zhou, C.; Ye, N. G-NAS: Generalizable neural architecture search for single domain generalization object detection. Proc. AAAI Conf. Artif. Intell. 2024, Vol. 38, 5958–5966. [Google Scholar] [CrossRef]
- Lee, W.; Hong, D.; Lim, H.; Myung, H. Object-aware domain generalization for object detection. Proc. AAAI Conf. Artif. Intell. 2024, Vol. 38, 2947–2955. [Google Scholar] [CrossRef]
- Liu, Y.; Zhou, S.; Liu, X.; Hao, C.; Fan, B.; Tian, J. Unbiased faster R-CNN for single-source domain generalized object detection. In Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2024; pp. 28838–28847. [Google Scholar]
- Danish, M.S.; Khan, M.H.; Munir, M.A.; Sarfraz, M.S.; Ali, M. Improving single domain-generalized object detection: A focus on diversification and alignment. In Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2024; pp. 17732–17742. [Google Scholar]
- Volpi, R.; Namkoong, H.; Sener, O.; Duchi, J.; Murino, V.; Savarese, S. Generalizing to unseen domains via adversarial data augmentation. Adv. Neural Inf. Process. Syst. 2018, 31. [Google Scholar]
- Li, L.; Gao, K.; Cao, J.; Huang, Z.; Weng, Y.; Mi, X.; Yu, Z.; Li, X.; Xia, B. Progressive domain expansion network for single domain generalization. In Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2021; pp. 224–233. [Google Scholar]
- Choi, S.; Das, D.; Choi, S.; Yang, S.; Park, H.; Yun, S. Progressive random convolutions for single domain generalization. In Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2023; pp. 10312–10322. [Google Scholar]
- Yuan, X.; Azimi, S.M.; Henry, C.; Gstaiger, V.; Codastefano, M.; Manalili, M.; Cairo, S.; Modugno, S.; Wieland, M.; Schneibel, A.; et al. Automated building segmentation and damage assessment from satellite images for disaster relief. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2021, XLIII-B3-2021, 741–748. [Google Scholar] [CrossRef]
- Mazzeo, A.; Renga, A.; Graziano, M.D. A systematic review of ship wake detection methods in satellite imagery. Remote Sens. 2024, 16, 3775. [Google Scholar] [CrossRef]
- Rabbi, J.; Ray, N.; Schubert, M.; Chowdhury, S.; Chao, D. Small-object detection in remote sensing images with end-to-end edge-enhanced GAN and object detector network. Remote Sens. 2020, 12, 1432. [Google Scholar] [CrossRef]
- Handan-Nader, C.; Ho, D.E. Deep learning to map concentrated animal feeding operations. Nat. Sustain. 2019, 2, 298–306. [Google Scholar] [CrossRef]
- Xu, X.; Yang, J.; Shi, W.; Ding, S.; Luo, L.; Liu, J. PhysAug: A physical-guided and frequency-based data augmentation for single-domain generalized object detection. Proc. AAAI Conf. Artif. Intell. 2025, Vol. 39, 21815–21823. [Google Scholar] [CrossRef]
- Jiang, H.; Yang, W.; Zhang, T.; Zhang, Y. Diffusion-based source-biased model for single domain generalized object detection. In Proceedings of the 2025 IEEE/CVF International Conference on Computer Vision (ICCV); IEEE, 2025; pp. 1548–1557. [Google Scholar] [CrossRef]
- Rao, Z.; Guo, J.; Tang, L.; Huang, Y.; Ding, X.; Guo, S. SRCD: Semantic reasoning with compound domains for single-domain generalized object detection. IEEE Trans. Neural Netw. Learn. Syst. 2025, 36, 12497–12506. [Google Scholar] [CrossRef] [PubMed]
- Xia, G.S.; Bai, X.; Ding, J.; Zhu, Z.; Belongie, S.; Luo, J.; Datcu, M.; Pelillo, M.; Zhang, L. DOTA: A large-scale dataset for object detection in aerial images. In Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2018; pp. 3974–3983. [Google Scholar]
- Michaelis, C.; Mitzkus, B.; Geirhos, R.; Rusak, E.; Bringmann, O.; Ecker, A.S.; Bethge, M.; Brendel, W. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv 2019, arXiv:1907.07484. [Google Scholar]
- Yang, J.; Shi, S.; Wang, Z.; Li, H.; Qi, X. ST3D: Self-training for unsupervised domain adaptation on 3D object detection. In Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2021; pp. 10363–10373. [Google Scholar]
- Li, Z.; Chen, Z.; Li, A.; Fang, L.; Jiang, Q.; Liu, X.; Jiang, J. Towards model generalization for monocular 3D object detection. arXiv 2022, arXiv:2205.11664. [Google Scholar]
- Qi, L.; Dong, P.; Xiong, T.; Xue, H.; Geng, X. DoubleAUG: Single-domain generalized object detector in urban via color perturbation and dual-style memory. ACM Trans. Multimed. Comput. Commun. Appl. 2024, 20, 126:1–126:20. [Google Scholar] [CrossRef]
- Park, C.; Jang, J.; Yoon, J.; Baek, M.; Paik, J. F-LGAM: Enhancing single domain generalized object detection through Fourier-based local and global amplitude mixup. In Proceedings of the 2025 IEEE International Conference on Image Processing (ICIP); IEEE, 2025; pp. 1474–1479. [Google Scholar] [CrossRef]
- Xu, M.; Qin, L.; Chen, W.; Pu, S.; Zhang, L. Multi-view adversarial discriminator: Mine the non-causal factors for object detection in unseen domains. In Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2023; pp. 8103–8112. [Google Scholar]
- Lu, C.; Su, S.; Li, H. Statistical-driven adaptive data augmentation for single-domain generalized object detection. Comput. Vis. Image Underst. 2026, 265, 104687. [Google Scholar] [CrossRef]
- Bi, Q.; Zhou, B.; Yi, J.; Ji, W.; Zhan, H.; Xia, G.S. GOOD: Towards domain generalized oriented object detection. ISPRS J. Photogramm. Remote Sens. 2025, 223, 207–220. [Google Scholar] [CrossRef]
- Wang, X.; Bin, K.; Wang, X.; Wan, P.; Hu, T.; Zhong, P. Adversarial spectral perturbation for single-domain generalized object detection. Pattern Recognit. 2026, 171, 112264. [Google Scholar] [CrossRef]
- Sun, Y.; Cao, B.; Zhu, P.; Hu, Q. Drone-based RGB-infrared cross-modality vehicle detection via uncertainty-aware learning. IEEE Trans. Circuits Syst. Video Technol. 2022, 32, 6700–6713. [Google Scholar] [CrossRef]
- Radford, A.; Kim, J.W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML). PMLR, 2021; pp. 8748–8763. [Google Scholar]
- Wang, K.; Fu, X.; Huang, Y.; Cao, C.; Shi, G.; Zha, Z.J. Generalized UAV object detection via frequency domain disentanglement. In Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2023; pp. 1064–1073. [Google Scholar]
- Han, J.; Wang, Y.; Chen, L. Style-adaptive detection transformer for single-source domain generalized object detection. Neurocomputing 2026, 685, 133583. [Google Scholar] [CrossRef]
- Hong, M.; Liu, F.; Gevaert, C.; Vosselman, G.; Cheng, H. Bridge: Basis-driven causal inference marries VFMs for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2026; pp. 31962–31973. [Google Scholar]
- Yuan, Y.; Tang, L.; Xu, Y.; Lin, C.; Chen, C.; Huang, Y.; Ding, X. Filling and disentanglement: Toward low- and high-order parallel single-domain generalization for SAR ship detection. IEEE Trans. Aerosp. Electron. Syst. 2025, 61, 3668–3682. [Google Scholar] [CrossRef]
- Lin, K.; Lin, Y.; Yuan, Y.; Tu, X.; Huang, Y.; Ding, X. Toward robust SAR ship detection: Domain-invariant learning via physics-driven augmentation and retrospective alignment. In Proceedings of the 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP); IEEE, 2026; pp. 20172–20176. [Google Scholar]
- Huang, H.; Lin, H.; Guo, J.; Huang, Y.; Ding, X.; Zeng, H. Single-domain generalization: A scattering characteristic-driven cross-attention model for oriented SAR object detection. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2026, 19, 12101–12115. [Google Scholar] [CrossRef]
- Al-Emadi, S.A.; Yang, Y.; Ofli, F. Benchmarking object detectors under real-world distribution shifts in satellite imagery. In Proceedings of the 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE, 2025; pp. 8299–8309. [Google Scholar] [CrossRef]
- He, H.; Ding, J.; Xu, B.; Xia, G.S. On the robustness of object detection models on aerial images. IEEE Trans. Geosci. Remote Sens. 2025, 63, 1–12. [Google Scholar] [CrossRef]
- Cordts, M.; Omran, M.; Ramos, S.; Rehfeld, T.; Enzweiler, M.; Benenson, R.; Franke, U.; Roth, S.; Schiele, B. The Cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016; pp. 3213–3223. [Google Scholar]
Table 1.
Target-domain mPC (averaged mAP@0.5, %) of the source-only and reference detectors under natural domain shift.
Table 1.
Target-domain mPC (averaged mAP@0.5, %) of the source-only and reference detectors under natural domain shift.
| Metric | ||
|---|---|---|
| mPC on DWD | 26.2 | 41.1 |
| mPC on RWDS-CZ | 8.8 | 16.4 |
Table 2.
Closed-gap values under natural domain shift. CG values are expressed as percentages. denotes DWD minus RWDS-CZ, in percentage points.
Table 2.
Closed-gap values under natural domain shift. CG values are expressed as percentages. denotes DWD minus RWDS-CZ, in percentage points.
| Method | DWD | RWDS-CZ | (pp) |
|---|---|---|---|
| OA-Mix | 10.9 | 6.7 | |
| SDAPM+DHDPG | 8.7 | ||
| GLT | 10.1 | 9.3 | |
| SCG | 12.8 | ||
| TBSA | 11.1 | ||
| PhysAug | 29.7 | 12.7 | |
| CP | 4.2 | ||
| Div | 34.2 | 11.3 | |
| F-LGAM | 9.9 | 7.3 |
Table 3.
Target-domain mPC (averaged mAP@0.5, %) of the source-only and reference detectors under synthetic common-corruption shift.
Table 3.
Target-domain mPC (averaged mAP@0.5, %) of the source-only and reference detectors under synthetic common-corruption shift.
| Metric | ||
|---|---|---|
| mPC on Cityscapes-C | 24.8 | 60.7 |
| mPC on DOTA-C | 37.6 | 54.3 |
Table 4.
Closed-gap values under synthetic common-corruption shift. CG values are expressed as percentages. denotes Cityscapes-C minus DOTA-C, in percentage points. SDAPM+DHDPG and Div were excluded because their augmentation sets overlap with the target corruptions.
Table 4.
Closed-gap values under synthetic common-corruption shift. CG values are expressed as percentages. denotes Cityscapes-C minus DOTA-C, in percentage points. SDAPM+DHDPG and Div were excluded because their augmentation sets overlap with the target corruptions.
| Method | Cityscapes-C | DOTA-C | (pp) |
|---|---|---|---|
| OA-Mix | 12.1 | 12.2 | |
| GLT | 26.9 | 31.5 | |
| SCG | 8.5 | 15.6 | |
| TBSA | 14.6 | 17.0 | |
| PhysAug | 25.8 | 33.4 | |
| CP | |||
| F-LGAM | 18.4 | 15.0 |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.