D-GAP: Improving Out-of-Domain Robustness via Dataset-Agnostic and Gradient-Guided Augmentation in Frequency and Pixel Spaces

Ruoqi Wang; Haitao Wang; Shaojie Guo; Qiong Luo

cs.CVNov 14, 2025

D-GAP: Improving Out-of-Domain Robustness via Dataset-Agnostic and Gradient-Guided Augmentation in Frequency and Pixel Spaces

Authors: Ruoqi Wang, Haitao Wang, Shaojie Guo, Qiong Luo

Organizations: HKUST(GZ) · SYSU · ECNU · HKUST

Abstract

Out-of-domain (OOD) robustness is challenging to achieve in real-world computer vision, especially in unsupervised domain adaptation scenarios, where shifts in image background, style, and acquisition instruments often degrade model performance. Generic augmentations show inconsistent gains under such shifts, whereas dataset-specific augmentations require expert knowledge and prior analysis. Moreover, prior studies show that neural networks adapt poorly to domain shifts because they exhibit a learning bias to domain-specific frequency components. Perturbing frequency values can mitigate such bias but overlooks pixel-level details, leading to suboptimal performance. To address these limitations, we propose D-GAP, a Dataset-agnostic and Gradient-guided augmentation method for the Amplitude spectrum (in frequency space) and the Pixel values. Unlike conventional handcrafted augmentations, D-GAP computes sensitivity maps in the frequency space from task gradients, which reflect how strongly the deep models respond to different frequency components, and uses the maps to adaptively interpolate amplitudes between source and target samples. We further propose a dual-space augmentation that jointly controls spectral bias and spatial fidelity by introducing a complementary pixel-space blending branch. This way, D-GAP turns augmentation from fixed, random, or manually designed perturbation into a model-response-adaptive intervention. Extensive experimental results show that the proposed method consistently outperforms both generic and dataset-specific domain adaptation methods, improving average OOD performance by +5.3% on four real-world datasets and +1.9% on three benchmark datasets. Code is available at https://github.com/RapidsAtHKUST/D-GAP.

Figures & tables

Figure 1 : Average amplitude sensitivity maps of ResNet50 trained across domains.

Figure 2 : Feature decomposition and augmentation examples. Top: For each dataset (iWildCam, Camelyon17, BirdCalls, Galaxy10), we show a source image, its corresponding augmented image generated by our method. Bottom: Based on the decomposition framework [ 49 ] , we annotate representative features in the datasets across $x_{\text{obj}}$ , $x_{d\text{:robust}}$ , $x_{d\text{:spu}}$ , $x_{\text{noise}}$ . Our method effectively randomizes $x_{d\text{:spu}}$ , varies $x_{d\text{:robust}}$ while preserving or $x_{\text{obj}}$ .

Figure 3 : Amplitude-only and phase-only reconstructions of two images. The amplitudes of Image A and B are first mixed, and the mixed amplitude is combined with the phase of Image A to generate an augmentation.

Aug	$\alpha$	$\beta$	$\gamma$	$\alpha/\gamma$	$\beta/\gamma$
w/o	0.007	0.008	0.021	0.33	0.38
Ours	0.212	0.030	0.051	4.16	0.59

Table 1 : Connectivity values on iWildCam.

Figure 4 : Overview of the Gradient-guided Amplitude Mix procedure. Given a source image $x_{1}$ and a target-domain image $x_{2}$ , we compute the Sensitivity Map $G(u,v)$ and the Mixing Map $D(u,v)$ . The mixed amplitude is then combined with the phase of $x_{1}$ .

Figure 5 : We plot the in-domain (ID) versus OOD performance across datasets (mean and std). ‘CL’ refers to ‘Connect Later’. ‘CP’ refers to ‘Copy Paste’. ‘SCJ’ refers to ‘Stain Color Jitter’.

Methods	PACS					OfficeHome
Methods	Art	Cartoon	Photo	Sketch	Avg.	Art	Clipart	Product	Real	Avg.
DeepAll	84.94	76.98	97.64	76.75	84.08	57.88	52.72	73.50	74.80	64.72
FACT (2023) [ 66 ]	89.68	81.06	97.02	83.76	87.88	60.34	54.85	74.48	76.55	66.56
SAM (2023) [ 65 ]	85.25	82.02	96.75	85.04	87.27	60.79	55.47	74.37	76.37	66.75
Su et al. (2024) [ 52 ]	79.88	78.03	95.93	73.15	81.75	64.13	54.83	75.60	75.36	67.71
AGST (2025) [ 62 ]	83.50	80.83	96.49	79.57	85.35	62.50	53.87	74.57	75.70	66.66

Table 2 : Model accuracy of leave-one-domain-out evaluation on PACS and OfficeHome.

Methods	MNIST	MNIST-M	SVHN	SYN	Avg.
DeepAll	95.8	58.8	61.7	78.6	73.7
FACT (2023) [ 66 ]	97.9	65.6	72.4	90.3	81.5
SAM (2023) [ 65 ]	97.8	66.8	73.2	90.7	82.1
Su et al. (2024) [ 52 ]	97.3	65.9	74.5	92.5	82.6
AGST (2025) [ 62 ]	97.3	63.1	70.2	93.2	81.0
Ours	98.3	68.5	77.2	93.8	84.5

Table 3 : Model accuracy of leave-one-domain-out evaluation on Digits-DG.

	iWildCam (F1)		Camelyon17 (Acc)		BirdCall (F1)		Galaxy10 (Acc)
	ID	OOD	ID	OOD	ID	OOD	ID	OOD
Ours w/o Augmentations	46.4	31.2	92.3	91.4	73.1	30.1	93.6	54.3
Pixel-only	45.4	29.1	87.5	69.8	75.2	32.4	95.5	68.2
Frequency-only	52.0	35.6	98.3	95.7	73.4	37.6	95.1	77.9
Mask-low	51.2	36.0	98.3	96.1	75.5	38.4	95.1	80.5
Mask-high	51.5	35.4	97.9	95.4	75.1	36.4	95.0	79.5

Table 4 : Ablation results across four real-world datasets.

Figure 6 : Performance of ConvNeXt and ViT on Galaxy10.

Method	PACS		Digits		OfficeHome
Method	ConvNeXt	ViT	ConvNeXt	ViT	ConvNeXt	ViT
SAM (2023) [ 65 ]	85.3	85.4	88.0	85.1	73.7	63.8
Su et al. (2024) [ 52 ]	87.9	75.3	88.9	85.6	73.9	64.6
AGST (2025) [ 62 ]	87.4	79.0	88.5	85.8	73.7	64.4
Ours	88.7	86.4	89.8	86.8	75.0	66.0

Table 5 : Average OOD performance on PACS, Digits-DG, and OfficeHome with ConvNeXt and ViT.

	Galaxy10		OfficeHome
High-gradient region	SC-CD	DC-SD	SC-CD	DC-SD
Top 10%	0.5449	0.5551	0.2774	0.3035
Top 20%	0.5485	0.5734	0.3332	0.3554
Top 30%	0.5497	0.5847	0.4372	0.4558

Table 6 : Jaccard similarity of top-ranked high-gradient frequency regions. SC-CD denotes same-class cross-domain pairs, while DC-SD denotes different-class same-domain pairs.

Camelyon17						iWildCam
Augmentations	$\alpha$	$\beta$	$\gamma$	$\alpha/\gamma$	$\beta/\gamma$	Augmentations	$\alpha$	$\beta$	$\gamma$	$\alpha/\gamma$	$\beta/\gamma$
w/o Augmentations	0.015	0.189	0.002	7.50	94.5	w/o Augmentations	0.007	0.008	0.021	0.33	0.38
RandAugment	0.181	0.186	0.114	1.59	1.63	RandAugment	0.224	0.125	0.130	1.72	0.96
Stain Color Jitter	0.029	0.116	0.004	7.25	29.0	Copy-Paste	0.254	0.031	0.063	4.03	0.49
Ours	0.120	0.073	0.003	40.0	24.3	Ours	0.212	0.030	0.051	4.16	0.59

Table 7 : Connectivity values across class, domain, and both on Camelyon17 and iWildCam datasets.

Method	Time / Iter. (ms)	Throughput (img/s)	Peak GPU Mem. (GB)
ERM	$24.46\pm 0.17$	$1308.3\pm 9.1$	0.86
MixUp	$24.71\pm 0.16$	$1294.9\pm 8.1$	0.88
DANN	$48.62\pm 0.24$	$658.2\pm 3.2$	1.55
LISA	$66.28\pm 0.19$	$482.8\pm 1.3$	0.88
SAM	$48.00\pm 0.30$	$666.6\pm 4.2$	0.91
FACT	$63.00\pm 0.31$	$507.9\pm 2.5$	1.94

Table 8 : Computational cost comparison under the same setting.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	MNIST	MNIST-M	SVHN	SYN	Avg.
DeepAll	95.8	58.8	61.7	78.6	73.7
FACT (2023) [ 66 ]	97.9	65.6	72.4	90.3	81.5
SAM (2023) [ 65 ]	97.8	66.8	73.2	90.7	82.1
Su et al. (2024) [ 52 ]	97.3 $\pm$ 0.33	65.9 $\pm$ 0.51	74.5 $\pm$ 0.48	92.5 $\pm$ 0.31	82.6
AGST (2025) [ 62 ]	97.3 $\pm$ 0.14	63.1 $\pm$ 0.45	70.2 $\pm$ 0.78	93.2 $\pm$ 0.37	81.0
Ours	98.3 $\pm$ 0.12	68.5 $\pm$ 0.26	77.2 $\pm$ 0.53	93.8 $\pm$ 0.24	84.5

Appendix

Table 9 : Model accuracy of leave-one-domain-out evaluation on Digits-DG (mean and standard deviation).

Methods	Art	Clipart	Product	Real	Avg.
DeepAll	57.88	52.72	73.50	74.80	64.72
FACT (2023) [ 66 ]	60.34	54.85	74.48	76.55	66.56
SAM (2023) [ 65 ]	60.79	55.47	74.37	76.37	66.75
Su et al. (2024) [ 52 ]	64.13 $\pm$ 0.50	54.83 $\pm$ 0.49	75.60 $\pm$ 0.61	75.36 $\pm$ 1.04	67.71
AGST (2025) [ 62 ]	62.50 $\pm$ 0.61	53.87 $\pm$ 0.29	74.57 $\pm$ 0.48	75.70 $\pm$ 0.78	66.66
Ours	64.82 $\pm$ 0.55	55.21 $\pm$ 0.37	77.76 $\pm$ 0.43	83.08 $\pm$ 0.25	70.22

Appendix

Table 10 : Model accuracy of leave-one-domain-out evaluation on OfficeHome (mean and std).

Methods	Art	Cartoon	Photo	Sketch	Avg.
DeepAll	84.94	76.98	97.64	76.75	84.08
FACT (2023) [ 66 ]	89.68	81.06	97.02	83.76	87.88
SAM (2023) [ 65 ]	85.25	82.02	96.75	85.04	87.27
Su et al. (2024) [ 52 ]	79.88 $\pm$ 0.53	78.03 $\pm$ 0.57	95.93 $\pm$ 0.21	73.15 $\pm$ 0.39	81.75
AGST (2025) [ 62 ]	83.50 $\pm$ 0.76	80.83 $\pm$ 0.48	96.49 $\pm$ 0.23	79.57 $\pm$ 0.62	85.35
Ours	89.34 $\pm$ 0.34	82.81 $\pm$ 0.63	97.99 $\pm$ 0.14	85.98 $\pm$ 0.19	89.03

Appendix

Table 11 : Model accuracy of leave-one-domain-out evaluation on PACS (mean and std).

Figure 7 : Effect of the frequency-window side-length ratio

r

on Galaxy10 and BirdCalls datasets.

Figure 8 : Effect of the mixing ratios

\lambda_{1}

and

\lambda_{2}

on Galaxy10 and BirdCalls datasets.

Method	Hyperparameters
ERM	Learning rate $=10^{-4}$ Weight decay $=0$
LISA	Learning rate $=10^{-4}$ Weight decay $=0$ Transform probability $=0.9$
MixUp	Learning rate $=10^{-4}$ Weight decay $=0$ Transform probability $=0.9$ $\alpha=0.4$
CutMix	Learning rate $=10^{-4}$ Weight decay $=0$ Transform probability $=0.9$ $\alpha=0.4$
Cutout	Learning rate $=10^{-4}$ Weight decay $=0$ Transform probability $=0.9$ $\alpha=0.4$
Randaugment	Learning rate $=10^{-4}$ Transform probability $=0.9$ Version = $Original$

Appendix

Table 12: Hyperparameter settings for baseline methods on Galaxy10.

Explore similar work

Domain-Aware Pruning: Sparsity and Domain Generalization via Regularized Probabilistic Masking
Aug 9, 2026Parham Sazdar, Mostafa Tavassolipour, Reshad HosseiniMultimodal Domain Generalization Sparsity
Department of Electrical and Computer Engineering, University of Tehran, Tehran, Iran
CPCANet: Deep Unfolding Common Principal Component Analysis for Domain Generalization
May 6, 2026Yu-Hsi Chen, Abd-Krim SeghouaneMultimodal Domain Generalization Subspace
The University of Melbourne
Spectral Gradient Surgery for Domain-Generalizable Dataset Distillation
May 13, 2026Minyoung Oh, Najeong Chae, Jae-Young SimDiffusion-Based Dataset Distillation Multimodal Domain Generalization
Graduate School of Artificial Intelligence Ulsan National Institute of Science and Technology (UNIST) Ulsan, Republic of Korea

Aug 9, 2026, cs.LGJ/K move · Enter open · S save

Domain-Aware Pruning: Sparsity and Domain Generalization via Regularized Probabilistic Masking

Parham Sazdar, Mostafa Tavassolipour, Reshad Hosseini

Department of Electrical and Computer Engineering, University of Tehran, Tehran, Iran

Multimodal Domain Generalization Sparsity Regularization

Open paper AskPDF arXiv

Domain generalization (DG) and neural network pruning are conventionally treated as distinct objectives, targeting out-of-distribution (OOD) robustness and model efficiency, respectively. In this work, we bridge this gap by introducing Domain-Aware Pruning (DAP), a framework that leverages network sparsity as a mechanism to implicitly enhance generalization to unseen domains. Diverging from standard binary mask optimization, DAP learns a continuous parameter retention probability

p \in [0, 1]

, framing network compression as a continuous probabilistic masking problem. By introducing a regularization objective that actively penalizes the retention of domain-sensitive weights during the mask training, DAP identifies a domain-invariant subnetwork. Empirical results across five DG benchmark datasets demonstrate that DAP achieves significant sparsity while consistently matching or exceeding the OOD performance of its dense counterparts. Crucially, DAP is an algorithm-agnostic framework that integrates seamlessly with existing DG pipelines without necessitating post-hoc fine-tuning. Beyond efficiency and generalization, we show that DAP natively provides increased robustness to adversarial perturbations and yields highly interpretable models, where the retained weights reliably encapsulate the most domain-invariant and task-critical representations.

Abstract

Explore similar work

Domain-Aware Pruning: Sparsity and Domain Generalization via Regularized Probabilistic Masking

CPCANet: Deep Unfolding Common Principal Component Analysis for Domain Generalization

Spectral Gradient Surgery for Domain-Generalizable Dataset Distillation