Single-Domain Generalized Object Detection

• Single-DGOD is defined as training an object detector on one labeled domain and evaluating it on diverse unseen target domains while preserving both classification and localization.
• Methods leverage techniques like semantic augmentation, contrastive learning, and transformer-based region modeling to handle appearance and structural shifts.
• Empirical results show improved mAP under weather-induced and corruption challenges, though issues remain with low-SNR categories and tiny objects.

Searching arXiv for recent and foundational papers on Single-Domain Generalized Object Detection to ground the article. {}

Single-Domain Generalized Object Detection (Single-DGOD), also written S-DGOD or SDGOD, is the problem of training an object detector on a single labeled source domain and evaluating it on multiple unseen target domains, with no target-domain images available during training. It differs from unsupervised domain adaptation, which assumes access to unlabeled target images to reduce the domain gap, and from multi-source domain generalization, which trains on multiple distinct source domains to learn domain-invariant features. The setting is unusually demanding because a detector must generalize from one training distribution to diverse and unpredictable test-time distributions shaped by weather, illumination, rendering, style, and sensor changes, while preserving both classification and localization accuracy (Han et al., 29 Apr 2025, He et al., 6 Feb 2025, Vidit et al., 2023).

1. Definition, scope, and detection-specific difficulty

Single-DGOD addresses a practical source-only regime in which the training set contains only one labeled domain, whereas deployment may involve night, rain, fog, artistic rendering, compression, blur, or other distribution shifts. In formal terms used by the literature, training uses a source dataset such as $D_s = \{(x_s^i, y_s^i, b_s^i)\}_{i=1}^{N_s}$ or $D_s = \{(X_s^i, B_s^i, Y_s^i)\}_{i=1}^{n_s}$, while testing occurs on unseen target samples $D_t = \{x_t^i\}_{i=1}^{N_t}$ or $D_t = \{X_t^i\}_{i=1}^{n_t}$ that share the same label space but differ in appearance and scene statistics (He et al., 6 Feb 2025, Li et al., 2024).

The detection setting is harder than single-domain classification because it must preserve instance semantics, object localization, and background discrimination simultaneously. Several papers argue that classification-oriented DG recipes can damage object features, invalidate spatial annotations, or ignore the special role of background in detection. OA-DG explicitly frames this as an object-aware problem, while SRCD emphasizes pseudo attribute-label correlation and the need to maintain semantic structural information among instances rather than merely compressing a feature space into a domain-invariant representation (Lee et al., 2023, Rao et al., 2023).

The literature also repeatedly distinguishes between nuisance shifts that are primarily stylistic and shifts that are more structural. Weather, illumination, and rendering are often treated as low-level appearance changes; density or resolution changes alter the observability of objects; and causal or semantic-context shifts induce spurious correlations between objects and backgrounds. This suggests that Single-DGOD has evolved not as one method family, but as a cluster of related attempts to reconcile diversification, invariance, localization, and semantic preservation under an extreme source-only constraint (Han et al., 29 Apr 2025, He et al., 6 Feb 2025, Li et al., 22 Oct 2025).

2. Formalization, metrics, and benchmark protocols

A standard objective is source-domain empirical risk minimization for object detection, augmented by DG-specific regularizers or augmentation schemes. In many papers, the base detector is Faster R-CNN, although DINO, GLIP, RT-DETR, YOLOv8-S, Mask R-CNN, DiffusionDet, and DINOv2-based variants also appear. Evaluation typically reports mean Average Precision at IoU 0.5, written as AP50 or mAP50, and computed across categories as

$\mathrm{mAP} = \frac{1}{c}\sum_i AP_i,$

where $AP_i$ is the area under the precision–recall curve for category $i$ (Han et al., 29 Apr 2025).

For corruption benchmarks, the literature also uses mean performance under corruption:

$$\mathrm{mPC} = \frac{1}{N_C}\sum_{C=1}^{N_C}\left(\frac{1}{N_S}\sum_{S=1}^{N_S} P_{C,S}\right),$$

where $P_{C,S}$ is the mAP under corruption $C$ at severity $D_s = \{(X_s^i, B_s^i, Y_s^i)\}_{i=1}^{n_s}$0. On Cityscapes-C, $D_s = \{(X_s^i, B_s^i, Y_s^i)\}_{i=1}^{n_s}$1 and $D_s = \{(X_s^i, B_s^i, Y_s^i)\}_{i=1}^{n_s}$2; on DWD, $D_s = \{(X_s^i, B_s^i, Y_s^i)\}_{i=1}^{n_s}$3 and $D_s = \{(X_s^i, B_s^i, Y_s^i)\}_{i=1}^{n_s}$4 (Xu et al., 2024).

Across papers, the dominant 2D urban-driving protocol trains only on a clear daytime source domain and evaluates on adverse or shifted domains such as Daytime-Foggy, Dusk-Rainy, Night-Clear or Night-Sunny, and Night-Rainy. One commonly used setup reports 19,395 training images and 8,313 in-domain test images for Daytime-Clear, with target domains Daytime-Foggy (3,775), Dusk-Rainy (3,501), Night-Clear (26,158), and Night-Rainy (2,494), over seven categories: person, car, bike, rider, motor, bus, truck (Han et al., 29 Apr 2025). Closely related benchmark variants use Daytime Sunny as source and Night Sunny, Dusk Rainy, Night Rainy, and Daytime Foggy as targets (Li et al., 2024).

Beyond urban driving, the literature also evaluates real-to-art generalization, typically training on PASCAL VOC and testing on Clipart1k, Watercolor2k, and Comic2k, and corruption robustness on Cityscapes-C. A 3D counterpart exists for point-cloud detection, where the same source-only DG principle is applied to LiDAR domain gaps driven by beam density, weather, and sensor geometry (Danish et al., 2024, Li et al., 2023).

3. Core methodological families

The field suggests several recurring design patterns: source-domain diversification, object-aware consistency, vision-language semantic prompting, style transfer, and explicit balancing of diversity and invariance. Representative examples are summarized below.

Family	Representative mechanism	Representative papers
Semantic augmentation with VLM priors	CLIP-based feature-space augmentation and text-based classification loss	"CLIP the Gap: A Single Domain Generalization Approach for Object Detection" (Vidit et al., 2023)
Compound-domain semantic reasoning	texture-based self-augmentation and local-global semantic reasoning	"SRCD: Semantic Reasoning with Compound Domains for Single-Domain Generalized Object Detection" (Rao et al., 2023)
Object-aware augmentation and contrast	OA-Mix and OA-Loss with foreground/background-aware contrastive learning	"Object-Aware Domain Generalization for Object Detection" (Lee et al., 2023)
Diversification plus consistency alignment	curated corruptions with multi-view classification and localization alignment	"Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment" (Danish et al., 2024)
Physical and frequency-based augmentation	atmospheric-optics-guided perturbation and frequency-domain illumination/occlusion synthesis	"PhysAug: A Physical-guided and Frequency-based Data Augmentation for Single-Domain Generalized Object Detection" (Xu et al., 2024)
Diversity–invariance balancing	domain-specific residual features with adaptive alignment weighting	"Single-Domain Generalized Object Detection by Balancing Domain Diversity and Invariance" (He et al., 6 Feb 2025)

The earliest line of work emphasized source diversification. CLIP the Gap transferred domain concepts such as weather and time of day into the detector through semantic augmentation in CLIP space and a text-based classification loss. SRCD instead constructed compound domains by texture-based self-augmentation in the Fourier domain and then modeled local and global semantic relations across ROI features. OA-DG pushed this logic further by introducing OA-Mix for multi-level object-aware augmentation and OA-Loss for contrastive and consistency learning that explicitly treats foreground and background differently (Vidit et al., 2023, Rao et al., 2023, Lee et al., 2023).

A second line argued that not all invariance is beneficial. DIDM states that excessive emphasis on invariance can cause the model to overlook the actual differences between images, complicate optimization, and discard valuable information. Its Diversity Learning Module preserves diversity in domain-specific features while limiting category semantics, and its Weighted Aligning Module modulates alignment strength according to output consistency, so that alignment is strong only when it materially helps detection (He et al., 6 Feb 2025).

A third line focused on augmentation quality rather than augmentation quantity. DivAlign showed that a carefully selected set of ImageNet-C and Fourier-based corruptions, followed by multi-view detection alignment for both class confidence and box localization, could outperform more complicated prior methods while also improving calibration. PhysAug made a related argument from atmospheric optics, replacing heuristic corruptions with a universal perturbation model that combines global non-uniform illumination, local particle-induced occlusions, and an atmospheric veil, all without changing network architecture or loss (Danish et al., 2024, Xu et al., 2024).

4. Region-aware, transformer-based, and generative extensions

Later work increasingly moved from global augmentation toward region-, instance-, and object-level modeling. PGST introduced phrase grounding-based style transfer on top of GLIP, learning target-domain style parameters $D_s = \{(X_s^i, B_s^i, Y_s^i)\}_{i=1}^{n_s}$5 from textual prompts without using target images and applying style transfer only to low-level features in order to preserve high-level semantics. Prompt-Driven Dynamic Object-Centric Learning, also based on scene prompts, introduced an object-centric gating module and a dynamic selective module that produce spatial and channel gating masks to emphasize object-centric features conditioned on prompt embeddings (Li et al., 2024, Li et al., 2024).

Transformer-based Single-DGOD then became explicit. SA-DETR built on DINO with a ResNet-50 backbone and introduced two modules: a Domain Style Adapter that uses channel-wise means and variances, EMA style bases, Wasserstein distance, and AdaIN-style rectification; and an Object-aware Contrastive Learning module that inserts class queries into the transformer encoder and constrains cross-attention with category masks derived from bounding boxes. In parallel, CRFI and CPRM reframed VLM guidance as region-aware interaction rather than global prompting, using fine-grained text-region InfoNCE objectives and cross-domain proposal refining and mixing to improve localization robustness (Han et al., 29 Apr 2025, Xu et al., 27 Apr 2025).

Several methods generalized style modeling beyond a single prompt or a single detector family. LDDS transferred style semantics from CLIP prompts into a UNet-based StyleNet, then applied image-level Fourier amplitude mixing and feature-level statistic mixing in a double-pipeline design that was explicitly described as model-agnostic across one-stage, two-stage, and transformer-based detectors. SE-COT argued that one-step prompts are weak for compound styles such as rain plus night, and therefore evolved style along a three-level chain of thought from keywords to phrases to sentences, using AdaIN on disentangled style features and a Class-Specific Prototype module for content supervision (Qin et al., 12 May 2025, Zhang et al., 13 Mar 2025).

Another branch emphasized causal or generative correction. Cauvis used cross-attention visual prompts and a dual-branch adapter that separates causal and spurious features while emphasizing high-frequency contours and edges. GoDiff used a latent diffusion model to generate pseudo-target images with instance-level control and then applied Cross-Style Instance Normalization and covariance matching to broaden the style distribution seen during training. FOUND added probabilistic Fourier augmentation and class-specific von Mises-Fisher regularization on normalized RoI features, arguing that amplitude predominantly encodes style whereas phase preserves structure (Li et al., 22 Oct 2025, Li et al., 2024, Wang et al., 13 Nov 2025).

The newest papers increasingly frame Single-DGOD as a feature-correction problem rather than an augmentation-coverage problem. LTFE simulated continuous distribution dynamics through progressive blur and noise, LSTM-based temporal modeling, and Neural ODE-generated dynamic kernels. CD-FKD used a frozen teacher on clean source images and a student on downscaled and corrupted views, distilling both global backbone features and instance-wise RoI features. MR-DCoT went furthest in this direction by treating unknown-domain generalization as manifold regression: Dual Chain-of-Thought modules generate structured off-manifold hard examples, and Class-Specific Prototype Anchoring learns a rectification operator that projects deviant features toward a source semantic manifold (Zhang et al., 13 Nov 2025, Lee et al., 17 Mar 2026, Zhang et al., 8 Jul 2026).

5. Empirical performance and comparative patterns

On the standard weather-driving benchmark, early source-only baselines left a large gap under adverse conditions. In CLIP the Gap, Faster R-CNN reported 34.4 on Night-Clear, 26.0 on Dusk-Rainy, 12.4 on Night-Rainy, and 32.0 on Day-Foggy, whereas Single-DGOD improved these to 36.6, 28.2, 16.6, and 33.5. The proposed CLIP-based method then reached 36.9, 32.3, 18.7, and 38.5, indicating that semantic augmentation and text-aligned classification were especially beneficial for fog and rain (Vidit et al., 2023).

Subsequent augmentation-and-alignment methods produced another step change. DivAlign reported 42.5 on Night Clear, 38.1 on Dusk Rainy, 24.1 on Night Rainy, and 37.2 on Daytime Foggy. PGST, using GLIP phrase grounding and object-level style transfer, reported 47.9 on Night Sunny, 44.5 on Dusk Rainy, 28.4 on Night Rainy, and 42.5 on Daytime Foggy, while also raising source-domain performance to 63.7 (Danish et al., 2024, Li et al., 2024).

The diversity–invariance and transformer lines then split the design space rather than converging to one template. DIDM reported an average of 34.0 over five domains versus 27.7 for Faster R-CNN, with 42.0 on night-sunny, 35.4 on dusk-rainy, 19.2 on night-rainy, and 39.3 on daytime-foggy. SA-DETR reported the best overall mAP50 in its ablation setting with source 64.8, Foggy 42.6, Dusk-Rainy 45.4, Night-Rainy 23.0, and Night-Clear 46.0, and improved the best prior result on Dusk-Rainy by 12.2 points (He et al., 6 Feb 2025, Han et al., 29 Apr 2025).

Corruption robustness on Cityscapes-C became another important reference point. PhysAug improved mPC from 15.4 to 22.6 on Cityscapes-C and from 30.2 to 37.5 on DWD without altering network architecture or loss. CRFI plus CPRM later reported 24.2 mPC on Cityscapes-C versus the same 15.4 baseline and 38.1 mPC on DWD versus 30.2. Cauvis raised Cityscapes-C mPC to 35.6, compared with 21.8 for OA-DG and 29.3 for FR+DINOv2, indicating that causal prompting plus high-frequency adaptation can substantially outperform earlier augmentation-only schemes in this benchmark family (Xu et al., 2024, Xu et al., 27 Apr 2025, Li et al., 22 Oct 2025).

Recent methods extended the ceiling again. LTFE reported, with ResNet-101, 43.1 on Night Sunny, 39.7 on Dusk Rainy, 24.3 on Night Rainy, and 41.2 on Day Foggy, and with Swin-T 53.1, 46.5, 33.4, and 46.4. CD-FKD reported an average target-domain [email protected] of 38.3, compared with 27.2 for Faster R-CNN and 35.5 for DivAlign in its comparison. MR-DCoT reported 47.4, 42.2, 28.1, and 42.8 with ResNet-101, and 54.6, 53.2, 36.4, and 48.3 with Swin-T, while also extending the same generalization logic to zero-shot semantic segmentation (Zhang et al., 13 Nov 2025, Lee et al., 17 Mar 2026, Zhang et al., 8 Jul 2026).

6. Limitations, misconceptions, and future directions

A recurring misconception in the literature is that stronger augmentation alone solves Single-DGOD. Several papers reject that view directly. SA-DETR argues that data augmentation methods are only effective when the augmented sample distribution approximates or covers the unseen scenarios; MR-DCoT similarly argues that finite simulations can hardly cover the dynamic variations of real-world scenarios. This suggests that augmentation is necessary but often insufficient, and helps explain the shift toward instance-level alignment, prototype anchoring, and feature rectification (Han et al., 29 Apr 2025, Zhang et al., 8 Jul 2026).

A second misconception is that invariance should always be maximized. DIDM explicitly states that excessive emphasis on invariance can cause the model to overlook the actual differences between images, complicate the training process, and lead to a loss of valuable information. Prompt-sensitive methods make a related point in a different language: PGST reports that weather-unrelated prompts reduce performance, and that over-stylization or applying style transfer to deeper layers can erode semantics. In other words, the problem is not simply learning domain-invariant features, but learning invariance in the correct subspace (He et al., 6 Feb 2025, Li et al., 2024).

Many current methods also retain specific modeling assumptions. SA-DETR relies on channel-wise mean and variance as style descriptors; GoDiff depends on pseudo-target quality and CLIP-RBF filtering of generated instances; PhysAug is designed for non-ideal imaging conditions arising from atmospheric optics and does not directly model geometry, viewpoint, or semantic appearance changes outside optics. These are not minor caveats: they specify where each method expects the dominant source of domain shift to lie (Han et al., 29 Apr 2025, Li et al., 2024, Xu et al., 2024).

Persistent failure modes remain visible at the category level. SA-DETR reports that Night-Rainy motor remains challenging, with 0.7 AP, and DIDM reports Motor at 3.8 AP under night-rainy conditions. Such results indicate that low-SNR categories, tiny objects, and extreme visibility loss remain unresolved even when overall mAP improves markedly (Han et al., 29 Apr 2025, He et al., 6 Feb 2025).

Future directions named in the literature are increasingly explicit. They include extending style-base methods to multi-source DG and test-time adaptation, using stronger or learned augmentation policies, combining DG with open-vocabulary detection, adopting adaptive or sample-conditioned $D_s = \{(X_s^i, B_s^i, Y_s^i)\}_{i=1}^{n_s}$6 and mixture-of-vMFs, improving text-domain style descriptions, learning style distributions rather than point estimates, and dynamically adjusting prototype neighborhoods or rectification radii (Han et al., 29 Apr 2025, Li et al., 2024, Wang et al., 13 Nov 2025, Zhang et al., 8 Jul 2026). A plausible implication is that the field is moving from simulation-driven robustness toward hybrid systems that combine controlled perturbation, semantic prompting, and explicit feature correction.

Single-DGOD has therefore developed from a narrow source-only detection problem into a broader research program on how to preserve object semantics under unknown domain shift. The progression from augmentation and feature alignment to object-aware reasoning, phrase grounding, transformer-based style adaptation, causal prompting, diffusion-based pseudo-target generation, temporal evolution, and manifold regression indicates that no single inductive bias dominates universally. Instead, the literature increasingly treats generalization as a structured detection problem in which style, geometry, objectness, and semantics must be managed jointly rather than independently (Vidit et al., 2023, Lee et al., 2023, Han et al., 29 Apr 2025).