acceptodds
Withdrawn from ICLR 2027

GS-Mix: Simple Pixel-level Mixing for Single-Domain Generalized Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pixel-level data augmentation has proven to be an effective strategy for single-domain generalized object detection (S-DGOD). However, existing methods struggle to balance augmentation intensity with the preservation of object-relevant structures, thereby restricting strong perturbations for diverse coverage. To address this, we analyze the key factors and influence mechanism in S-DGOD domain shift, and accordingly propose a vicinal distribution that synthesizes samples with domain shift perturbations (noise and texture) while robustly preserving structure. To instantiate this distribution, we construct a simple yet effective pixel-level intra-domain mixing framework, termed Global-context Structure-texture Mix (GS-Mix). Specifically, global-context mixing (GCM) yields diverse noise perturbations through linear interpolation between source images, and structure-texture mixing (STM) uses structure masks to apply strong texture transformations. These modules cover global-to-local strong augmentations and avoid destroying structural information. GS-Mix can be easily taken as a plug-and-play data augmentation module and introduces almost no extra costs in the training phase. Experimental results on real-world domain generalization benchmarks, including adverse weather, underwater, and medical imaging, show the superior performance and easy scalability of our GS-Mix compared to existing methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Object Detection, Data Augmentation, Generalization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7zWCOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

GS-Mix is a pixel-level augmentation framework for single-domain generalized object detection. It combines global-context MixUp (GCM) with edge-guided structure–texture mixing (STM), motivated by an object-centric vicinal risk formulation. The method aims to diversify noise and texture while retaining structural cues. Validation covers adverse weather, real-to-cartoon, medical, and underwater benchmarks, with experiments across three detector families, component ablations, and efficiency analysis.

Critical strengths

1.Substantial empirical gains: On YOLOv8s, GS-Mix improves mPC from 29.9 to 37.1 on adverse weather and from 34.9 to 45.3 on real-to-cartoon.

2.Practical simplicity: The augmentation requires no changes to detector architectures or losses and introduces no additional inference cost.

3.Broad validation: Experiments span diverse domain shifts and detector architectures. Ablations show complementary benefits from GCM and STM, with particularly strong STM gains under appearance changes.

Critical weaknesses

1.The claims in lines 58-67 regarding the limitations of feature-level and frequency-domain augmentation lack sufficient supporting evidence or analysis. These broad characterizations make the motivation for the proposed approach less convincing.

2.The main contribution appears to be another variant of MixUp: GCM adopts standard image interpolation, while STM introduces an edge-based mixing mask with conventional texture augmentations. The conceptual advance over existing mixing strategies is not sufficiently established.

3.Prior S-DGOD work usually evaluates on both DWD and Cityscapes-C. The absence of Cityscapes-C experiments limits validation of the paper’s claimed robustness to common corruptions.

4.mPC is used in the main results but defined only in Appendix A.1. Its meaning and calculation should be explained at first use in the main text to make the results accessible to a broader readership.

Questions for authors

1.What specific evidence supports the claims in lines 58–67 about semantic disruption in feature-level augmentation and restricted perturbation directions in frequency-domain methods? How does this evidence establish the need for GS-Mix?

2.What evidence isolates the benefit of STM’s structural mask from its conventional texture augmentations and mixing strength? Tables 7 and 9 establish the benefits of STM and its texture augmentations, but do not isolate the importance of aligning the mixing mask with image structure. Is there evidence that the edge-based mask outperforms a random or spatially shuffled mask with matched coverage and mixing weights, under the same augmentation settings?

3.Was Cityscapes-C evaluated under a comparable source-only training protocol? If results are available, please report them, as they would directly inform the assessment of robustness to common corruptions. Otherwise, please explain its omission.

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

This paper proposes Global-context Structure-texture Mix (GS-Mix), a pixel-level augmentation method for single-domain generalized object detection. Global-context mixing (GCM) linearly blends two source images, while structure-texture mixing (STM) uses a soft edge mask to combine one image with a color-transformed and filtered image. Annotations from both images are retained for training. The proposed method randomly selects between the two branches during training and is motivated by an object-centric vicinal risk minimization formulation. No additional trainable modules, auxiliary losses, or inference operations are required. Experiments demonstrate effectiveness of the proposed method on adverse-weather and real-to-artistic benchmarks with Faster R-CNN and YOLOv8 backbones, with further evaluations using RT-DETR on medical and underwater datasets.

Critical strengths

GS-Mix offers a detector-agnostic pixel-level augmentation strategy without introducing additional trainable modules, specialized auxiliary losses, or inference-time operations. The reported improvements across Faster R-CNN, YOLOv8, and RT-DETR support its usefulness across different detector backbones. Experiments cover several types of domain shift and include component ablations, augmentation-operator analysis, parameter sensitivity studies, and qualitative results. These results demonstrate the effectiveness of the proposed method.

Critical weaknesses

(1) The evidence to support how the proposed structure-preservation mechanism works is limited. The component ablations show that GCM and STM improve performance. However, they do not isolate how much of this gain comes from preserving object structure, as opposed to image mixing, appearance changes, or interactions with existing augmentations. These results therefore establish the method effectiveness more clearly than they explain its mechanism. Moreover, the OC-VRM equations describe the augmentation distribution and training objective, but do not establish that the actual mixing operations preserve object structure. The characterization of classical VRM as typically isotropic also requires a more careful justification.

(2) The comparison includes several augmentation-based S-DGOD methods, but omits recent related methods such as UICL (AAAI 2026), LTFE (AAAI 2026), and AdvSP (Pattern Recognition 2026).

(3) The main tables draw on results from different papers without clearly indicating whether their training settings are comparable. For example, the Faster R-CNN baseline achieves 28.6 mPC on Diverse Weather in this submission, compared with 29.4 in DiffDet and 26.2 in OA-DG. It should be clarified whether the results are copied from prior works or reproduced by the authors. The differences in pretraining, standard augmentations, and training schedules should also be clarified.

Questions for authors

(1) What are the evidences showing that STM preserves the object structure beyond improving detection performance? Can the authors quantify object visibility, boundary or shape preservation, and the proportion of annotations that still correspond to recognizable objects after mixing? These measurements would help assess whether the structure-preservation claim is valid. (2) Is Fig. 2 a conceptual illustration? It does not specify how the features were extracted or projected. Could the authors provide analysis of real object features from the trained detector, using a common extraction and projection setup for original source images, augmented images, and unseen target images? (3) The underwater training set pools images from domains 1-6. Does this protocol fit the single-domain setting? If these are distinct source domains, should this experiment be presented as a supplementary multi-source evaluation?

Reviewer_5wXnOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies single-domain generalized object detection and proposes GS-Mix, a pixel-level data augmentation framework. Motivated by an object-centric vicinal risk minimization view, the method perturbs noise and texture while preserving object structure, through two modules: global-context mixing, which blends source images, and structure-texture mixing, which combines one image’s edge structure with another’s augmented texture. The approach is plug-and-play with negligible cost, and is validated on adverse weather, real-to-cartoon, underwater and medical benchmarks across several detectors, with ablations and comparisons to other augmentation methods.

Critical strengths

S1. GS-Mix is a plug-and-play, pixel-level augmentation that requires no architectural changes or auxiliary models, adds no inference cost, and incurs little training overhead, making it easy to adopt across different detectors.

S2. The method is evaluated on diverse generalization settings, from adverse weather and artistic styles to underwater and medical imaging, and consistently improves the baselines, enabling lightweight detectors to perform competitively with heavier methods.

Critical weaknesses

W1. Limited novelty and insufficient positioning against prior work.

As acknowledged in Sec. 3.1, GCM applies Mixup with merged detection annotations; its linear blending operation matches Ultralytics’ detection Mixup. Its gains therefore demonstrate the effectiveness of an established augmentation. STM’s motivation—reducing texture bias through structure-preserving appearance perturbations—also has substantial precedent. However, the seminal texture-bias analysis [1] is not discussed, and [2] is cited only to motivate domain shift, without addressing its stylized-training approach to improving detector robustness. Closely related texture-randomization methods for single-domain generalization, including RandConv [3] and Pro-RandConv [4], are neither discussed nor compared. Although the component ablations demonstrate improvements beyond GCM, these omissions leave STM’s methodological distinction and empirical advantage over relevant prior augmentations insufficiently established.

W2. The claimed mechanism is not supported.

The connection between the OC-VRM formulation (Eqs. 1–3) and the proposed augmentation operators is insufficiently established. The factors η\eta, tt, and uu are introduced through examples, the composition gg remains unspecified, and no operational criterion is provided to verify that the operators vary noise or texture while preserving structure. It is therefore unclear what testable constraints the formulation imposes on the design.

STM is described as transferring texture onto xix_i’s structure, yet Eq. (7) inserts xjax_j^a outside xix_i’s edge regions and retains yjy_j. This can introduce xjx_j’s own objects and spatial layout, making the interpretation of texture variation at fixed structure questionable. The reported ablations demonstrate performance gains but do not establish that these gains arise from reduced texture bias.

[1] ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness, ICLR’19
[2] Benchmarking robustness in object detection: Autonomous driving when winter is coming, 2019
[3] Robust and Generalizable Visual Representation Learning via Random Convolutions, ICLR’21
[4] Progressive Random Convolutions for Single Domain Generalization, CVPR’23

Questions for authors

Q1. How does GS-Mix compare with established texture-debiasing augmentations, e.g., stylized training [2], RandConv [3] and Pro-RandConv [4], at least on the real-to-cartoon and adverse-weather tasks? This determines whether STM offers an advantage over existing approaches to the same problem.

Q2. Could the authors ablate STM by (i) taking texture from an augmented copy of xix_i itself instead of xjx_j, and (ii) using y~=yi\tilde{y}=y_i only? This would show whether the gains indeed come from the claimed structure–texture decoupling.

Q3. Could the authors compare full STM against (i) applying Ψ(xi)\Psi(x_i) directly, and (ii) same-image STM, m(xi)⊙xi+(1−m(xi))⊙Ψ(xi)m(x_i)\odot x_i+(1-m(x_i))\odot\Psi(x_i), using yiy_i in both cases? These controls would help assess the contributions of appearance augmentation, edge-guided mixing, and the additional content and supervision introduced by xjx_j.

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

Related papers

Discussion (0)

Sign in to comment.