SRasP: Self-Reorientation Adversarial Style Perturbation for Cross-Domain Few-Shot Learning

Wenqian Li; Pengfei Fang; Hui Xue

cs.CVMar 5, 2026

SRasP: Self-Reorientation Adversarial Style Perturbation for Cross-Domain Few-Shot Learning

Authors: Wenqian Li, Pengfei Fang, Hui Xue

Organizations: the School of Computer Science and Engineering, Southeast University, Nanjing 210096, China · the Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, Nanjing 211189, China

Abstract

Cross-Domain Few-Shot Learning (CD-FSL) aims to transfer knowledge from a seen source domain to unseen target domains, serving as a key benchmark for evaluating the robustness and transferability of models. Existing style-based perturbation methods mitigate domain shift but often suffer from gradient instability and convergence to sharp minima.To address these limitations, we propose a novel crop-global style perturbation network, termed Self-Reorientation Adversarial Style Perturbation (SRasP). Specifically, SRasP leverages global semantic guidance to identify incoherent crops, followed by reorienting and aggregating the style gradients of these crops with the global style gradients within one image. Furthermore, we propose a novel multi-objective optimization function to maximize visual discrepancy while enforcing semantic consistency among global, crop, and adversarial features. Applying the stabilized perturbations during training encourages convergence toward flatter and more transferable solutions, improving generalization to unseen domains. Extensive experiments are conducted on multiple CD-FSL benchmarks, demonstrating consistent improvements over state-of-the-art methods.

Figures & tables

Fig. 1: (a) Given an input image, multiple local crops are extracted. The training process applies one of three crop selection strategies, including random-crop selection, concept-crop selection and incoherent-crop selection, each leading to different gradient directions during optimization. (b) The gradient cosine similarity across training epochs shows that the proposed method maintains consistently higher stability compared with other perturbation methods, indicating a more reliable update trajectory. (c) Loss-surface visualizations further demonstrate that our incoherent-crop perturbation drives the model toward flatter and more generalizable minima.

Fig. 2: Comparison between Vanilla Global Perturbation and our Self-Reorientation Perturbation. Left : In the vanilla approach, a global style perturbation is directly applied to the feature map of the entire input. Right : In our Self-Reorientation Perturbation, the input image is first divided into incoherent crops. These crop style gradients are then reoriented and aggregated with the global style gradient.

Fig. 3: Overview of the proposed SRasP. SRasP first samples multiple localized crops from an input image and identifies incoherent regions that exhibit semantic inconsistency with the global content. Style gradients are extracted from both crop-level and global features, where incoherent crop gradients are reoriented toward the global semantic direction and aggregated through a self-reorientation gradient ensemble to resolve gradient conflicts. The resulting stable and semantically guided global style gradient is then used to synthesize hard yet meaningful adversarial style perturbations via AdaIN. In addition, a consistency–discrepancy triplet objective is employed to maximize visual style diversity while preserving semantic alignment among global, crop, and adversarial representations, enabling SRasP to generate stronger style variations and improve cross-domain generalization.

	Method	Venue	Back.	FT	ChestX	ISIC	EuroSAT	CropDisease	CUB	Cars	Places	Plantae	Aver.
1-shot	GNN [ 39 ]	ICLR’18	RN10	✗	22.00 $\pm$ 0.46	32.02 $\pm$ 0.66	63.69 $\pm$ 1.03	64.48 $\pm$ 1.08	45.69 $\pm$ 0.68	31.79 $\pm$ 0.51	53.10 $\pm$ 0.80	35.60 $\pm$ 0.56	43.55
	FWT [ 13 ]	ICLR’20	RN10	✗	22.04 $\pm$ 0.44	31.58 $\pm$ 0.67	62.36 $\pm$ 1.05	66.36 $\pm$ 1.04	47.47 $\pm$ 0.75	31.61 $\pm$ 0.53	55.77 $\pm$ 0.79	35.95 $\pm$ 0.58	44.14
	ATA [ 40 ]	IJCAI’21	RN10	✗	22.10 $\pm$ 0.20	33.21 $\pm$ 0.40	61.35 $\pm$ 0.50	67.47 $\pm$ 0.50	45.00 $\pm$ 0.50	33.61 $\pm$ 0.40	53.57 $\pm$ 0.50	34.42 $\pm$ 0.40	43.84
	StyleAdv [ 11 ]	CVPR’23	RN10	✗	22.64 $\pm$ 0.35	33.96 $\pm$ 0.57	70.94 $\pm$ 0.82	74.13 $\pm$ 0.78	48.49 $\pm$ 0.72	34.64 $\pm$ 0.57	58.58 $\pm$ 0.83	41.13 $\pm$ 0.67	48.06
	FLoR [ 41 ]	CVPR’24	RN10	✗	23.11 $\pm$ 0.31	38.11 $\pm$ 0.51	62.90 $\pm$ 0.65	73.64 $\pm$ 0.64	49.99 $\pm$ 0.62	37.41 $\pm$ 0.54	53.18 $\pm$ 0.79	40.10 $\pm$ 0.64	47.31
	SVasP [ 12 ]	AAAI’25	RN10	✗	23.23 $\pm$ 0.35	37.63 $\pm$ 0.58	72.30 $\pm$ 0.82	75.87 $\pm$ 0.73	49.49 $\pm$ 0.72	35.27 $\pm$ 0.57	59.07 $\pm$ 0.81	41.22 $\pm$ 0.62	49.26

TABLE I: Quantitative comparison to state-of-the-arts methods on eight target datasets based on ResNet-10, which is pretrained on miniImageNet. Accuracy of 5-way 1-shot/5-shot tasks with 95 confidence interval are reported. “ FT ” with ✓ means finetuning is used, vice versa. “ Aver. ” means “Average Accuracy” of the eight datasets. The optimal results are marked in bold .

Method

Venue

Back.

ChestX

ISIC

EuroSAT

CropDisease

CUB

Cars

Places

Plantae

Aver.

1-shot

StyleAdv [ 11 ]

CVPR’23

ViT-S

✗

22.92

\pm

0.32

33.05

\pm

0.44

72.15

\pm

0.65

81.22

\pm

0.61

84.01

\pm

0.58

40.48

\pm

0.57

72.64

\pm

0.67

55.52

\pm

0.66

57.75

FLoR [ 41 ]

CVPR’24

ViT-S

✗

22.78

\pm

0.31

34.20

\pm

0.45

72.39

\pm

0.67

81.81

\pm

0.60

84.60

\pm

0.57

40.71

\pm

0.58

73.85

\pm

0.65

51.93

\pm

0.65

57.78

CD-CLS [ 45 ]

NeurIPS’24

ViT-S

✗

22.93

\pm

0.33

34.21

\pm

0.45

74.08

\pm

0.63

83.51

\pm

0.60

82.90

\pm

0.60

41.67

\pm

0.57

69.17

\pm

0.66

51.88

\pm

0.67

57.54

SVasP [ 12 ]

AAAI’25

ViT-S

✗

22.68

\pm

0.30

34.49

\pm

0.46

72.50

\pm

0.62

80.82

\pm

0.62

85.56

\pm

0.57

40.51

\pm

0.59

75.93

\pm

0.66

56.25

\pm

0.65

58.59

REAP [ 46 ]

ICML’25

ViT-S

✗

23.62

\pm

0.31

37.21

\pm

0.44

74.69

\pm

0.60

84.04

\pm

0.59

82.09

\pm

0.62

42.23

\pm

0.60

72.02

\pm

0.66

55.56

\pm

0.67

58.93

ReCIT [ 26 ]

ICML’25

ViT-S

✗

23.27

\pm

0.31

35.13

\pm

0.44

74.56

\pm

0.60

84.76

\pm

0.59

82.30

\pm

0.58

42.89

\pm

0.58

72.64

\pm

0.66

55.60

\pm

0.65

58.89

TABLE II: Quantitative comparison to state-of-the-arts methods on eight target datasets based on ViT-small, which is pretrained on ImageNet1K by DINO. Accuracy of 5-way 1-shot/5-shot tasks with 95 confidence interval are reported.

	SR	$\mathcal{L}_{\mathrm{CDTO}}$	$\mathcal{L}_{\mathrm{con}}$	ChestX	ISIC	EuroSAT	CropDisease	CUB	Cars	Places	Plantae	Aver.
1-shot	-	-	-	22.61 $\pm$ 0.35	33.99 $\pm$ 0.56	69.04 $\pm$ 0.84	70.25 $\pm$ 0.77	45.73 $\pm$ 0.68	32.77 $\pm$ 0.56	53.91 $\pm$ 0.78	36.19 $\pm$ 0.62	45.56
	✓			23.29 $\pm$ 0.35	34.99 $\pm$ 0.56	70.67 $\pm$ 0.81	73.85 $\pm$ 0.78	45.75 $\pm$ 0.68	33.73 $\pm$ 0.56	54.55 $\pm$ 0.77	38.70 $\pm$ 0.59	46.94
	✓	✓		22.62 $\pm$ 0.34	36.23 $\pm$ 0.57	70.17 $\pm$ 0.83	72.24 $\pm$ 0.77	46.08 $\pm$ 0.68	34.56 $\pm$ 0.56	55.46 $\pm$ 0.77	39.81 $\pm$ 0.62	47.15
	✓		✓	22.92 $\pm$ 0.36	36.77 $\pm$ 0.53	72.40 $\pm$ 0.84	75.60 $\pm$ 0.78	48.43 $\pm$ 0.69	35.16 $\pm$ 0.59	57.91 $\pm$ 0.81	40.19 $\pm$ 0.62	48.67
	✓	✓	✓	23.39 $\pm$ 0.37	37.92 $\pm$ 0.60	74.90 $\pm$ 0.82	76.82 $\pm$ 0.77	50.62 $\pm$ 0.68	36.20 $\pm$ 0.60	60.11 $\pm$ 0.75	41.94 $\pm$ 0.61	50.24
5-shot	-	-	-	25.16 $\pm$ 0.35	47.46 $\pm$ 0.54	86.30 $\pm$ 0.54	92.07 $\pm$ 0.44	63.37 $\pm$ 0.68	49.26 $\pm$ 0.63	75.09 $\pm$ 0.64	57.66 $\pm$ 0.65	62.07

TABLE III: Ablation study of the proposed method with different component combinations. “SR” indicates Self-Reorientation Gradient Ensemble module.

	Mining Strategy	BSCD-FSL	mini -CUB	Aver.
1-shot	Baseline	48.97	42.15	45.56
	Concept	52.31	46.37	49.34
	Random	52.42	46.57	49.49
	Incoherent (Ours)	53.26	47.22	50.24
5-shot	Baseline	62.75	61.35	62.05
	Concept	65.28	65.21	65.24

TABLE IV: Performances on different crop mining strategies. The average accuracy (%) is reported.

Fig. 4: Meta-training loss curves of the baseline and the proposed SRasP.

	$\xi$	BSCD-FSL	mini -CUB	AVer.
1-shot	$\xi=0.01$	52.57	46.41	49.49
	$\xi=0.1$ (Ours)	53.26	47.22	50.24
	$\xi=1$	52.68	46.47	49.58
5-shot	$\xi=0.01$	64.93	64.51	64.72
	$\xi=0.1$ (Ours)	65.92	65.65	65.78
	$\xi=1$	65.11	64.62	64.86

TABLE V: Performances on different

\xi

. The average accuracy (%) is reported.

	$\lambda$	BSCD-FSL	mini -CUB	AVer.
1-shot	$\lambda=0$	52.77	46.61	49.69
	$\lambda=0.2$	53.26	47.22	50.24
	$\lambda=0.4$	52.91	46.64	49.78
	$\lambda=0.6$	52.49	46.36	49.43
	$\lambda=0.8$	52.30	45.88	49.09
	$\lambda=1$	51.99	45.67	48.83

TABLE VI: Performances on different

\lambda

. The average accuracy (%) is reported.

	$\kappa_{1}$ , $\kappa_{2}$	BSCD-FSL	mini -CUB	AVer.
1-shot	Same	52.43	46.18	49.30
1-shot	Different (Ours)	53.26	47.22	50.24
5-shot	Same	65.01	64.68	64.85
5-shot	Different (Ours)	65.92	65.65	65.78

TABLE VII: Performances on whether use same

\kappa_{1}

\kappa_{2}

. The average accuracy (%) is reported.

Fig. 5: Performances on different numbers of crops $k$ .

Fig. 6: Loss landscape visualization of different methods on the BSCD-FSL benchmark. SRasP produces markedly flatter loss landscapes across all target domains compared to the baseline and SVasP.

Fig. 7: Grad-CAM visualizations of different methods on four representative datasets. SRasP produces cleaner and more object-centric activation patterns across diverse target domains.

Explore similar work

Prompt-Driven Simulation with Feature Perturbation for Cross-Domain Few-Shot Object Detection
Aug 2, 2026Linhai Zhuo, Junxi Cai, Tianwen Qian +2Synthetic Data Augmentation Few-Shot Domain Adaptation
Fuzhou University · East China Normal University · Xiamen University +1
Reviving In-domain Fine-tuning Methods for Source-Free Cross-domain Few-shot Learning
May 12, 2026Yaze Zhao, Yicong Liu, Yixiong Zou +2Source-Free Domain Adaptation Few-Shot Domain Adaptation
Training-free Cross-domain Few-shot Segmentation via Robust Semantic Representation and Matching
Jun 23, 2026Sujun Sun, Mingwu Ren, Haofeng ZhangFew-Shot Learning Cross-Domain Transfer Learning
School of Computer Science and Engineering, Nanjing University of Science and Technology, China · State Key Laboratory of Intelligent Manufacturing of Advanced Construction Machinery, China

Aug 2, 2026J/K move · Enter open · S save

Prompt-Driven Simulation with Feature Perturbation for Cross-Domain Few-Shot Object Detection

Linhai Zhuo, Junxi Cai, Tianwen Qian, Qingping Zheng+1

Fuzhou University · East China Normal University · Xiamen University+1

Synthetic Data Augmentation Few-Shot Domain Adaptation Data Augmentation Few-Shot Object Detection Cross-Domain Generalization Object Detection

Open paper AskPDF arXiv

Data augmentation, which simulates diverse visual variations to expand the source distribution and induce synthetic domain shifts, is a simple yet effective strategy for mitigating severe domain shifts and limited labeled target data in cross-domain few-shot object detection (CD-FSOD). Existing approaches rely on conventional data augmentation, such as Color-Jitter, Mosaic, and background-centric adaptation (e.g., Domain-RAG), which are limited in modeling complex domain shifts and often lead to suboptimal performance. In this paper, we propose PSP-FSOD, a principled framework that integrates prompt-driven domain simulation with feature perturbation regularization to improve generalization in CD-FSOD. To enable controllable domain synthesis, we design a prompt-driven strategy that leverages the visual grounding capability of large VLMs to jointly model foreground and background variations, generating semantically consistent yet domain-diverse training samples. Moreover, we adopt a grounding-aware generation scheme that guides object placement and alleviates semantic-spatial misalignment, thereby improving foreground adaptation. To ensure training stability and robustness, we further introduce a noise-induced feature perturbation mechanism that injects Gaussian noise into multi-scale intermediate features with distribution correction, encouraging consistent predictions under perturbations and reducing reliance on domain-specific cues. Extensive experiments demonstrate that PSP-FSOD produces high-quality domain-diverse supervision and learns domain-invariant representations, consistently improving performance across CD-FSOD benchmarks.

Abstract

Explore similar work

Prompt-Driven Simulation with Feature Perturbation for Cross-Domain Few-Shot Object Detection

Reviving In-domain Fine-tuning Methods for Source-Free Cross-domain Few-shot Learning

Training-free Cross-domain Few-shot Segmentation via Robust Semantic Representation and Matching