WorksComputer Science
Rethinking Feature Perturbation for Single-Source Domain Generalization: A Representation-Space Analysis and Spectral Case Study
Abstract
Feature perturbation is widely used in single-source domain generalization (SSDG), yet it is typically evaluated by task performance rather than by the representation changes it induces. We investigate whether perturbations can be compared through measurable descriptors of these changes. We introduce a four-dimensional representation-space description comprising feature displacement , relational displacement , a task-region ratio , and subspace deviation . We use these descriptors as an empirical analysis framework, not as a universal geometry, to characterize existing perturbation mechanisms. Across two object-detection settings involving style and weather shifts, we observe distinct geometric profiles and a concentration of relatively strong out-of-domain (OOD) configurations within a bounded region of the descriptor space. These findings do not establish a universal optimum but motivate investigating whether regions associated with stronger SSDG performance can be identified without target-domain access. As a case study, we introduce Dual-Invariant Spectral Perturbation (DISP), a non-learned Fourier-phase perturbation with explicit controls over phase magnitude , channel coupling , and frequency support . We empirically examine how these controls affect the proposed descriptors. DISP achieves strong OOD performance in our experiments, illustrating how explicit perturbation controls can be analyzed through a representation-space description. We hope this viewpoint can contribute to more interpretable, controllable, and ultimately testable approaches to feature perturbation in SSDG.
Introduction
Deep networks can perform remarkably well when training and test data follow similar distributions, yet deployment often violates this assumption. Changes in appearance, sensing conditions, scene composition, or weather can substantially alter the visual distribution, as encountered in UAV inspection and other real-world vision applications (Benelmostafa et al.; Fan et al.; Oh et al.). This motivates domain generalization (DG), which seeks models that transfer to unseen target domains without target data during training (Gulrajani & Lopez-Paz). The single-source setting (SSDG) is particularly challenging because only one training environment is observed, forcing the model to construct useful variability without directly observing cross-domain representation changes (Qiao & Peng). In object detection, this challenge is further complicated by appearance, texture, scale, background context, localization, and partial visibility.
Among existing approaches, feature perturbation provides a direct way to introduce variability inside the network: an intermediate representation is replaced during training by
without requiring target-domain observations. However, perturbations are typically evaluated through downstream performance, commonly , which does not reveal what they change internally. Two perturbations can achieve similar detection scores while differing in feature displacement, relational structure, or spatial allocation; conversely, a larger feature change does not necessarily yield better OOD generalization.
We therefore ask:
Can feature perturbations be described in a common representation-space language that is informative beyond final detection accuracy?
To explore this question, we construct a four-dimensional empirical descriptor space capturing how far the representation moves, how much its channel relations change, where the displacement occurs relative to object regions, and whether it leaves a dominant clean-feature subspace. We use these descriptors to compare several perturbation mechanisms under a common protocol. Across two benchmarks, relatively stronger OOD configurations tend to occupy a bounded portion of this space. We treat this as a hypothesis rather than a universal optimum, motivating a broader question: does SSDG exhibit a reproducible region where source diversification and preservation of task-relevant structure are jointly achieved, and can it be identified using only source-domain information?
To make this question concrete, we introduce Dual-Invariant Spectral Perturbation (DISP) as a simple case study. DISP perturbs Fourier phase while preserving spectral amplitude, with three explicit intervention variables: phase magnitude , channel coupling , and spectral support . Rather than assuming that these controls directly determine the proposed descriptors, we measure how representation geometry responds to controlled changes in the operator, providing a simple setting in which perturbation design and representation analysis can be studied together.
Our contributions are intentionally exploratory: (i) a four-dimensional descriptor space for comparing feature perturbations; (ii) an empirical characterization of existing mechanisms and recurring descriptor patterns associated with stronger OOD performance across two benchmarks; (iii) DISP as a non-learned spectral case study with explicit, measurable intervention variables; and (iv) open research questions on whether descriptor spaces can become predictive, source-selectable, and useful for interpretable SSDG perturbation design.
Related Work
Feature perturbation methods create training variability by modifying intermediate representations, but they differ substantially in what is changed and how the perturbation is generated. We organize the literature around these mechanisms while keeping the central question of this work in view: how can the resulting representation change be described in a common language?
Stochastic and statistical perturbation. DSU (Li et al., b) models channel-wise feature statistics as random variables. Normalization Perturbation (NP) (Fan et al.) perturbs low-level statistics, with a particular emphasis on shallow CNN representations. MixStyle (Zhou et al.) mixes statistics across samples, while CrossNorm exchanges them (Tang et al.). Adversarial Style Augmentation (ASA) searches for challenging perturbations in statistic space (Zhang et al., a). These methods are attractive because their interventions are simple to implement, but their effect on the full representation is typically summarized through the statistics being manipulated rather than through a common representation-space description.
Distribution-level and cross-sample perturbation. EFDM extends statistical matching by transforming empirical feature distributions (Zhang et al., b). CCFP introduces cross-sample feature perturbation under semantic consistency constraints (Li et al., a), while Domain Feature Perturbation exploits domain information through an auxiliary classifier (Wang et al., a). Other SSDG work has explored combined image- and feature-level perturbations, such as CPerb and MixPatch (Zhao et al.). Recent object-detection studies also emphasize the balance between diversity and invariance. under single-source shifts ((Danish et al.; He et al.)). These approaches broaden the space of perturbation mechanisms, but they also make direct comparison of the induced feature changes more difficult.
Spectral perturbation. Fourier methods provide a different way of constructing domain variation. AmpMix perturbs Fourier amplitude while preserving phase ((Xu et al.)), motivated by the different roles of amplitude and phase in signal representation ((Oppenheim & Lim)). FIESTA introduces uncertainty-guided spectral manipulation ((Oh et al.)), while UniFreqSDG learns spectral perturbation ranges and strengths for SSDG ((Liu et al.)). More recently, AdvSP uses learnable spectral perturbations for single-domain generalized object detection ((Wang et al., b)). These studies support the usefulness of frequency-domain interventions, but they do not provide a shared descriptor language for comparing the representation changes produced by otherwise different mechanisms.
Position of this work. Our objective is therefore not to claim that existing perturbation methods lack control or interpretation. Several expose useful hyperparameters, and learned methods may have considerably more expressive internal control than their user-facing parameterization suggests. Instead, we ask whether their effects can be described under the same representation-space descriptors. We use this perspective to analyze existing mechanisms and then to study DISP as a simple, explicit intervention whose controls can be related to measurable representation changes.
Table 1: Exposed control levers of feature perturbation mechanisms. Learned indicates whether the perturbation itself is optimized jointly with the task. Control levers lists parameters exposed to the practitioner beyond a fixed application probability; most methods expose none, relying instead on a hardcoded scale, no free parameter, or values predicted internally.
| Method | Modified object | Learned | Control levers |
| MixStyle (Zhou et al.) - 2024 | channel mean / std | No | (mixing distribution) |
| CrossNorm (Tang et al.) - ICCV 2021 | channel mean / std | No | – |
| DSU (Li et al., b) - ICLR 2022 | channel mean / std | No | none (hardcoded scale) |
| NP (Fan et al., b) - ICLR 2023 | channel mean / std | No | none (hardcoded scale) |
| EFDM (Zhang et al., b) - CVPR 2022 | empirical distribution | No | none (parameter-free) |
| AmpMix (Xu et al.) - 2023 | Fourier amplitude | No | (mixing bound) |
| FIESTA (Oh et al.) - 2024 | Fourier amplitude | No | noise scale |
| CCFP (Li et al., a) - ICCV 2023 | cross-sample stat. shift | Yes | none (architecture only) |
| UniFreqSDG (Liu et al.) - ACM MM 2024 | amplitude, radius, gating | Yes | none (predicted internally) |
| AdvSP (Wang et al., b) - 2026 | Fourier spectrum | Yes | none (predicted internally) |
| DISP (ours) | Fourier phase | No | , , band |
Table 1. Exposed control levers of feature perturbation mechanisms. Learned indicates whether the perturbation itself is optimized jointly with the task. Control levers lists parameters exposed to the practitioner beyond a fixed application probability; most methods expose none, relying instead on a hardcoded scale, no free parameter, or values predicted internally.
A Representation-Space Descriptor View
Motivation
A perturbation is usually defined by the operation it applies: changing statistics, matching a distribution, exchanging information across samples, or modifying a spectrum. Such definitions describe the mechanism, but they do not necessarily reveal the resulting representation shift in a comparable form. We therefore introduce a descriptor view based on four complementary questions: how far does the representation move, how much does its relational structure change, where does the displacement occur with respect to object regions, and whether does the displacement leave a dominant clean-feature subspace? The resulting descriptor is
We emphasize that this construction is an analysis lens, not a claim that these four quantities form a complete or canonical geometry of representation space. Each descriptor captures one aspect of the perturbation, and each also depends on design choices that we make explicit below.
Descriptors
Let denote a clean feature map, its perturbed counterpart, and . Let denote their spatially flattened forms, and let denote a numerical stabilizer.
Feature displacement .
measures the relative size of the induced feature change. It is intentionally simple: the purpose is to provide a common magnitude reference rather than to assert that larger displacement is better or worse.
Relational displacement . We define the channel Gram matrix
and measure its change as
This descriptor captures changes in second-order inter-channel structure and complements : two perturbations can have similar displacement magnitude while modifying channel relations differently.
Task-region ratio . Let denote the union of projected ground-truth object regions on the feature grid and let . For a spatial mask , define
We then use
Values above one indicate proportionally larger displacement in background than in object regions. This quantity should be interpreted as a spatial allocation descriptor, not as a direct measure of task importance.
Subspace deviation . Let contain the top- right singular vectors of the clean , with . We define
and
Larger values indicate that a larger fraction of perturbation energy lies outside the chosen dominant clean-feature subspace. The reference subspace is always computed from , not , providing a common frame across methods. As with all descriptors, the choice of is a modeling choice; we use 0.2 throughout the present analysis.
What This Descriptor View Can and Cannot Establish
The descriptor view supports comparisons that are difficult to make from the perturbation operator alone. It allows us to ask whether methods with similar downstream performance induce similar representation changes, whether large changes are systematically associated with OOD performance, and whether object-region allocation or subspace deviation provides information not captured by a single magnitude measure. At the same time, the descriptors do not establish causality. A correlation between a descriptor and OOD performance can arise from the architecture, training dynamics, data composition, or interactions between the perturbation and the learned representation. Nor do we assume that the favorable region identified in our current experiments is universal. The intended contribution is to formulate a measurable object for future study: a representation descriptor space in which perturbations can be compared, analyzed, and eventually, perhaps, selected without target-domain supervision.
DISP: A Spectral Case Study
From Analysis to an Explicit Intervention
The descriptor view naturally suggests a second question: if perturbation effects can be measured in a common representation space, can we design an intervention whose mechanism exposes interpretable variables that can be systematically varied?
We study this question through Dual-Invariant Spectral Perturbation (DISP). DISP is not proposed as a solution to the descriptor-selection problem. Instead, it is a deliberately simple case study that makes the interaction between perturbation controls and representation descriptors explicit.
DISP Operator
Let be an intermediate feature map and its per-channel spatial Fourier transform,
DISP preserves amplitude and perturbs phase within a selected band :
where the shared field is common to channels and the independent field is sampled per channel, both from standard Gaussian noise. The perturbation is normalized to unit RMS in the selected band before being scaled by , and the DC component is excluded.
The controls have distinct mechanistic roles: sets the RMS magnitude of the phase perturbation, determines the balance between shared and channel-specific phase variation, and specifies the spectral support of the perturbation. Importantly, we do not assume that these variables independently determine the four descriptors. Their effect on representation geometry is measured empirically.
Analytical Properties
Energy preservation. Because DISP modifies phase while preserving spectral amplitude, under an orthonormal Fourier transform,
Consequently,
The first statement follows from equal feature energy and the triangle inequality; the second follows from the continuity of the inverse Fourier transform.
Relational invariance at full coupling. For the channel Gram matrix above, Parseval’s identity gives
When , every channel receives the same phase offset, so the relative phase between channel pairs is unchanged. Under the corresponding full-spectrum/Hermitian reconstruction,
For intermediate , we make no monotonicity claim: the resulting relational change depends on the existing phase structure and the sampled channel-specific component.
Frequency localization. The band specifies where the phase perturbation is applied. Fourier analysis motivates treating frequency localization as a meaningful intervention variable because different frequency components can contribute to different spatial scales (Oppenheim & Lim). We do not derive a deterministic mapping from low-, mid-, or high-frequency perturbations to or ; those effects are treated as empirical questions.
Experiments and Analysis
Experimental Setup
We use a fixed YOLO26-nano detector implemented in Ultralytics v8.4.49. All methods use the same detector, optimization protocol, image preprocessing, perturbation probability, and perturbation locations; only the perturbation mechanism changes. Motivated by the observation that low-level domain-sensitive statistics are strongly represented in shallow CNN layers (Fan et al., b), we insert perturbation modules at backbone layers 0, 1, and 3 (the transition). We evaluate two complementary settings: Pascal VOC 2007 as the source with unseen Clipart, Watercolor, and COCO targets, and clear-weather IDID as the source with unseen rain, fog, and snow conditions. Dataset composition and splits are given in Appendix § 12. We compare source-only training (Baseline) with CCFP, AmpMix, MixStyle, DSU, EFDM, FIESTA, UniFreqSDG, NPOriginal, and DISP. OOD performance is reported using with perturbations disabled during validation and inference. Geometric descriptors are computed on paired clean and perturbed source-domain evaluation features at .
Main Out-of-Domain Detection Results
Table 2 reports the source and unseen-target values obtained in our current runs. In these experiments, DISP has the highest reported mean OOD score among the compared configurations on both settings: 9.00% on the VOC family and 48.70% on IDID. The absolute margins over AmpMix are small (+0.47 and +0.33 points, respectively), so we do not interpret them as evidence of a large performance improvement. Rather, they show that the proposed spectral intervention is at least competitive under the present protocol. On VOC, DISP obtains the highest score among the compared methods on Clipart and Watercolor, while other methods remain competitive on COCO. On IDID, DISP obtains the highest reported Rain score and the highest source-domain score in our table, while DSU and FIESTA are stronger on Snow and Fog, respectively. This variation is useful for our main question because it illustrates that no single perturbation mechanism dominates all target conditions. Figure 1 shows representative detections. On Clipart, DISP recovers additional persons in one example where Baseline and AmpMix produce an incorrect aeroplane prediction. In other style-shift examples, competing methods give different responses to objects outside the benchmark label space; we interpret these cases only as evidence of different model responses, not as proof that one prediction is semantically correct. The fog example similarly illustrates that a perturbation useful under one weather condition need not preserve every anomaly detection under another condition. These observations motivate examining the representation-level behavior rather than relying on isolated qualitative outcomes.
| Method | VOC | Clipart | Watercolor | COCO | Mean OOD | IDID | IDID | IDID | IDID | Mean OOD |
| Baseline | 25.4 | 4.6 | 4.7 | 9.3 | 6.20 | 71.2 | 43.5 | 42.7 | 45.4 | 47.07 |
| DSU (ICLR 2022) | 25.0 | 4.7 | 8.7 | 9.3 | 7.57 | 68.1 | 47.1 | 49.8 | 40.7 | 45.87 |
| NPOriginal (ICLR 2023) | 25.2 | 5.1 | 6.5 | 9.0 | 6.87 | 68.8 | 45.7 | 47.9 | 40.4 | 44.67 |
| EFDM (CVPR 2022) | 24.0 | 5.7 | 8.1 | 8.4 | 7.40 | 62.9 | 40.7 | 50.2 | 42.4 | 44.43 |
| FIESTA | 26.5 | 6.1 | 7.7 | 9.5 | 7.77 | 71.2 | 40.5 | 51.9 | 45.9 | 46.10 |
| MixStyle | 25.7 | 6.9 | 7.4 | 9.6 | 7.97 | 64.7 | 44.5 | 48.7 | 43.2 | 45.47 |
| UniFreqSDG (ACM MM 2024) | 26.9 | 6.1 | 8.5 | 9.9 | 8.17 | 67.4 | 44.4 | 54.2 | 38.7 | 45.77 |
| CCFP (ICCV 2023) | 27.6 | 4.8 | 9.2 | 9.9 | 7.97 | 67.1 | 42.6 | 45.8 | 38.6 | 42.33 |
| AmpMix | 27.8 | 6.1 | 9.8 | 9.7 | 8.53 | 69.8 | 46.3 | 53.4 | 45.4 | 48.37 |
| DISP (Ours) | 27.8 | 6.7 | 10.7 | 9.6 | 9.00 | 72.2 | 45.1 | 55.8 | 45.2 | 48.70 |
Table 2.

Figure 1. Representative detections on unseen style and weather domains. The examples illustrate different responses of perturbation mechanisms under domain shift; they are intended as qualitative context rather than quantitative evidence.
Table 2: Single-source domain generalization results on Pascal VOC 2007 and IDID. Source performance (VOC and IDID) is evaluated on the training domain. Mean OOD target performance is computed over the unseen target domains only. Light gray denotes the best method (Top 1); Bold denotes the second-best (Top 2).
Descriptor Space and OOD Performance
The preceding table establishes the task-level outcomes. We next ask whether the methods also differ systematically in the representation descriptors introduced in § 3. We profile the same mechanisms at using paired clean and perturbed features.
Table 3 shows that the perturbation mechanisms occupy distinct regions of the descriptor space. EFDM is a clear outlier in the current measurements, with the largest , , and on both benchmarks. DSU has the largest reported on both settings. These observations by themselves say nothing about whether the corresponding coordinates are good or bad; they simply show that different perturbation rules induce measurably different representation changes.
| Method | Pascal VOC 2007 | IDID | ||||||
| AmpMix | 0.165 | 0.227 | 1.128 | 0.551 | 0.338 | 0.488 | 1.331 | 0.403 |
| CCFP | 0.127 | 0.188 | 1.108 | 0.549 | 0.344 | 0.503 | 1.298 | 0.421 |
| FIESTA | 0.132 | 0.191 | 1.107 | 0.552 | 0.312 | 0.459 | 1.306 | 0.415 |
| NPOriginal | 0.168 | 0.203 | 1.165 | 0.631 | 0.337 | 0.477 | 1.334 | 0.458 |
| UniFreqSDG | 0.133 | 0.193 | 1.122 | 0.532 | 0.330 | 0.469 | 1.311 | 0.398 |
| DSU | 0.197 | 0.325 | 1.069 | 0.684 | 0.348 | 0.528 | 1.277 | 0.489 |
| EFDM | 0.430 | 0.723 | 1.319 | 0.461 | 0.543 | 0.795 | 1.589 | 0.400 |
| MixStyle | 0.358 | 0.511 | 1.047 | 0.547 | 0.468 | 0.636 | 1.239 | 0.430 |
| DISP (Ours) | 0.159 | 0.198 | 1.143 | 0.575 | 0.325 | 0.463 | 1.312 | 0.452 |
Table 3. Representation descriptors measured at on the source-domain evaluation split. Values are means over valid samples under the protocol in Appendix § 11.
More interestingly, several of the methods with relatively strong OOD performance appear within a bounded portion of the four-dimensional descriptor space. Figure 2 visualizes this pattern. We call the region a favorable region only in the descriptive sense that the currently plotted higher-performing configurations tend to occur there. We do not define a universal boundary, fit a predictive classifier, or claim that the region is optimal. The appropriate interpretation is therefore a research hypothesis: perhaps SSDG perturbations that generalize well do not need to maximize any individual descriptor, but instead occupy a constrained combination of representation displacement, relational change, task-region allocation, and subspace exploration.

Figure 2. Proposed descriptor space on (a) Pascal VOC and (b) IDID. The shaded area is a visual guide highlighting a bounded region in which several relatively strong OOD-performing configurations occur. It is not an analytically defined optimum and should be interpreted as an empirical observation to be tested in broader settings.
This observation is deliberately weaker than a geometric law. With only two benchmarks, one architecture, and a small set of perturbation mechanisms, the current evidence cannot establish that the same region would persist under different backbones, tasks, or shifts. What the experiment does provide is a concrete empirical object that can be investigated further.
Controlled Exploration with DISP
We next use DISP to examine whether explicit perturbation variables produce systematic trajectories in the descriptor space. We sweep phase magnitude , channel coupling , and frequency support . For comparison, AmpMix is evaluated over . The full values are reported in Appendix § 14.§ 1. Figure 3 shows that varying produces the clearest progression in feature displacement: in the mid-band sweep, larger substantially increases . Changes in and affect combinations of relational and directional descriptors, but not necessarily in a monotonic one-coordinate fashion. This is consistent with the distinction made in § 4: the parameters have interpretable meanings at the operator level, while their representation-level consequences emerge from the interaction with the learned features. The reported VOC sweep also illustrates that stronger perturbation is not automatically associated with better OOD performance. Increasing from 0.1 to 0.9 increases from approximately 0.158 to 0.534 while mean OOD decreases from 13.7% to 10.4%. We do not interpret this as evidence for a causal relation between and OOD accuracy. It is instead a concrete example showing why perturbation strength and downstream performance should not be treated as interchangeable quantities.

Figure 3. Controlled exploration of the representation descriptor space on Pascal VOC. AmpMix varies along its amplitude-mixing parameter , while DISP generates distinct trajectories by changing phase magnitude , channel coupling , and frequency support . The shaded region is the same descriptive favorable-region guide used in Figure 2.
Taken together, these experiments suggest a useful division of labor: the perturbation mechanism provides explicit intervention variables, while the descriptor space provides a measurement layer for recording what those interventions actually do. Whether this measurement layer can eventually be used for source-only selection or optimization remains an open question.
Discussion: What Does This View Open Up?
The main outcome of this work is not a claim that one perturbation is universally superior, but that feature perturbation may benefit from being treated as an object of representation-level analysis. Our measurements provide several preliminary indications of this view: perturbation mechanisms exhibit distinct descriptor profiles even when applied at the same network locations; larger feature displacement does not consistently imply stronger OOD performance; relatively strong configurations tend to occupy a bounded region of the descriptor space; and DISP shows that a simple perturbation operator can expose intervention variables whose effects can be measured rather than inferred only from final accuracy. These observations motivate several questions that extend beyond the current numerical margin of DISP over the strongest baseline. First, does a reproducible favorable region exist across datasets, backbones, detectors, and domain shifts? A stronger evaluation should test whether high-performing perturbations consistently occupy similar regions after suitable normalization. Second, can such a region be identified without target data? Our current analysis uses target performance to characterize favorable configurations and therefore does not yet provide a source-only selection rule. Future SSDG protocols could instead exploit source-domain stability, cross-augmentation consistency, simulated validation domains, or related source-only signals. Third, which descriptors actually matter? Our four descriptors capture magnitude, relational change, spatial allocation, and subspace deviation, but they are neither unique nor necessarily complete; alternatives include CKA-like similarity, covariance geometry, spatial relations, class-conditional distances, attention structure, and task-gradient information. Finally, can descriptor space enable controllable perturbation design? DISP provides an initial step, but future work could formulate perturbation selection as a source-only optimization problem that targets a desired descriptor region while preserving source-task performance. In this view, feature perturbation could move beyond a collection of independent heuristics toward a more systematic design problem. More broadly, descriptor space may offer a useful intermediate vocabulary between intervention mechanism and task performance, provided that its geometry proves stable and informative across broader settings.
Conclusion
We revisited feature perturbation in SSDG from a representation-centric perspective. Rather than treating perturbation strength as the primary object of study, we introduced a compact descriptor space capturing feature displacement, relational change, task-region allocation, and subspace deviation. Across our experiments, perturbation mechanisms exhibit distinct descriptor profiles, with relatively stronger OOD-performing configurations tending to occupy a bounded portion of this space. DISP provides a simple case study of this perspective by perturbing Fourier phase while explicitly controlling magnitude, channel coupling, and frequency support. This highlights the distinction between intervention variables and the emergent representation descriptors they produce. The main contribution is therefore a change in viewpoint:
Measure Characterize Question Explore.
Rather than defining a universal optimal region, this work opens the question of whether such representation-space structure can be identified from source data and used to make feature perturbation more interpretable, controllable, and systematic.
AI Disclosure
In this work, we used generative AI tools (large language models) for research execution and methodology refinement, specifically to critically evaluate our scientific rationale, challenge methodological assumptions, and propose refinements to our experimental design. Additionally, we used AI tools to assist in implementing software code for our experimental pipelines and validation tests. We also used generative AI tools to edit and rephrase the manuscript text to improve overall readability, clarity, and language precision. All AI-assisted contributions were rigorously verified by the authors: LLM-generated code was thoroughly inspected, tested, and validated for execution correctness, and all conceptual feedback was critically evaluated through manual literature and theoretical review. We take full responsibility for the final content of this work, including all text, scientific claims, code artifacts, and experimental results.
DISP Algorithm
Algorithm 1 gives the forward pass. The implementation uses a real-valued FFT representation; the non-redundant coefficients determine the corresponding Hermitian-symmetric spectrum for reconstruction. The phase offset is normalized to unit RMS inside the selected band before scaling by , and the module is disabled at inference.
Algorithm 1 DISP forward pass
Require: Feature map ; ; ; frequency band ; probability
if not training or then
return
end if
; ;
Sample and
Construct radial mask and set the DC component to zero
Normalize to unit RMS on the selected band
return
The module is inserted as a drop-in replacement inside the standard convolution–normalization–activation block, with the perturbation applied between normalization and activation. It is bypassed during inference.
Mathematical Properties
The analytical statements below use full-spectrum notation. The implementation stores the non-redundant half-spectrum and reconstructs the Hermitian counterpart needed for real-valued outputs.
Proposition 1 (Energy invariance). DISP preserves the spectral amplitude at every channel and frequency and therefore satisfies
Proof. Under orthonormal Fourier normalization, Parseval’s identity gives and . Since , the two spectral norms are equal.
Corollary 1 (Bounded continuous feature displacement). For DISP,
Proof. The bound follows from the triangle inequality and Proposition 1. For the limit, for fixed , and continuity of the inverse Fourier transform gives .
Lemma 1 (Gram matrix in the frequency domain). For the channel Gram matrix in § 3,
Proof. Apply Parseval to the channel-wise inner product and substitute .
Proposition 2 (Relational invariance at full coupling). Under the full-spectrum/Hermitian reconstruction described above, at ,
Proof. At , every channel receives the same phase offset. Therefore for every channel pair, while amplitudes remain unchanged. Lemma 1 then gives .
For , the pairwise phase difference contains the channel-specific term
so the effect on depends on the existing representation and is not assumed to be monotonic. Likewise, the influence of on the directional descriptors is treated as an empirical question.
Geometric Measurement Protocol
Paired features. For a fixed source image, clean and perturbed features are obtained from otherwise identical forward passes with the perturbation disabled and enabled, respectively. The detector is kept in evaluation mode during profiling so that batch-normalization statistics remain fixed. For the Baseline, the perturbation is disabled in both passes, yielding and therefore .
Evaluation split. Geometric descriptors are computed on the source-domain evaluation split at the representation, using batches of 16. The same protocol is applied to the source domains of both benchmarks.
Per-image computation. With clean/perturbed features and ground-truth boxes, Algorithm 2 implements the four definitions in Section 3. We use throughout.
Algorithm 2 Per-image representation descriptors
Require: ; ground-truth boxes ; stride ;
Project the union of boxes to the feature grid to obtain ;
Compute from the top- right singular vectors of ,
;
return
Images without annotated objects are excluded from aggregate computation. The two directional descriptors are also excluded when falls below a fixed numerical tolerance. Reported geometric values are means over the remaining valid images.
Datasets and Domain Splits
Style-shift benchmark: Pascal VOC Clipart, Watercolor, COCO. The source domain is Pascal VOC 2007, comprising the standard 20-class label space
with 5,011 training images and 4,952 test images (VOC in Table 2), used as the sole training and in-domain evaluation data. Clipart and Watercolor originate from the cross-domain detection benchmark of Inoue et al.. Clipart shares the full 20-class VOC label space and contributes 999 test images; Watercolor covers only a 6-class subset, {person, bird, dog, cat, car, bicycle}, and contributes 1,997 test images, so mean OOD statistics involving Watercolor are computed only over these six shared classes, not the full 20. COCOComp denotes the COCO validation set filtered and remapped to the VOC label space, contributing 4,952 test images. No target-domain data is used during training in any configuration. Figure 4 (“Overview of Visual Style Divergence and Class Imbalance across PASCAL VOC, Clipart, Watercolor, and COCO”) shows representative images and the per-class instance distribution for all four domains.

Figure 4. Cross-Domain Object Detection Benchmark and Category Distribution. Visual samples with bounding box annotations (left) alongside class frequency distributions (right) across four benchmark domains: PASCAL VOC (source/real), Clipart (vector graphics), Watercolor (artistic style), and COCO (VOC-compatible split). The figure highlights both the substantial style/visual shift across domains and the intrinsic category imbalance within each dataset, illustrating the challenges of cross-domain object detection.
| Domain | Train | Test | Classes |
| VOC (source) | 5,011 | 4,952 | 20 |
| Clipart (target) | – | 999 | 20 |
| Watercolor (target) | – | 1,997 | 6 (shared subset) |
| COCO (target) | – | 4,952 | 20 (remapped) |
Table 4. Style-shift benchmark: image counts and label space by domain.
Weather-shift benchmark: IDID clear-weather rain, fog, snow. IDID (Aitelhaj et al.) is a UAV power-line inspection dataset comprising 1,600 images, split 70%/20%/10% into training/test/validation (approximately 1,120/320/160 images). The source domain (IDID in Table 2) consists of clear-weather images only; models are trained exclusively on this split. A single anomaly class is annotated, representing insulator shells that are either flashover-damaged or broken. The three target domains, , , and , apply simulated weather conditions to the corresponding test images, following the protocol of Benelmostafa et al., and are never seen during training. Figure 5 shows representative images from , , , and side by side, illustrating the visual severity of each condition.

Figure 5. Illustrative Images of Datasets: Sequential column display showing IDID-Original (first column), followed by IDID-Snow, IDID-Rain, and IDID-FOG (last column).(Benelmostafa et al.)
| Domain | Train | Test | Classes |
| (source) | 1 | ||
| (target) | – | 1 | |
| (target) | – | 1 | |
| (target) | – | 1 |
Table 5. Weather-shift benchmark: image counts and label space.
Implementation and Training Details
All experiments use Ultralytics v8.4.49 and YOLO26-nano under otherwise identical optimization settings. Pascal VOC experiments use inputs and 70 epochs; IDID uses inputs and 100 epochs without pretrained weights. Feature perturbations are applied stochastically with and disabled during evaluation. The training runs are tracked with ClearML on a single NVIDIA T4 GPU.
The same perturbation locations are used for all compared methods. For methods whose original formulation uses a different intervention level, the current study adopts a common feature-level implementation at the same backbone locations so that the comparison isolates the behavior of the perturbation mechanism under a shared detector/training pipeline. This should be understood as a level-matched experimental adaptation rather than a claim of exact reproduction of every original training recipe.

Figure 6a. (a) VOC:

Figure 6b. (b) VOC: box loss
fig:6: Representative training curves on the VOC benchmark.

Figure 7a. (a) IDID:

Figure 7b. (b) IDID: box loss
fig:7: Representative training curves on the IDID benchmark.
Extended Results
Controlled Exploration: DISP Sweep
| Config | VOC | Clipart | Water. | COCO | Mean | ||||
| (0.1, 0.5, mid) | 27.8 | 6.7 | 10.7 | 9.6 | 13.7 | 0.158 | 0.197 | 1.143 | 0.574 |
| (0.1, 0.1, mid) | 26.7 | 5.6 | 9.2 | 10.1 | 12.9 | 0.153 | 0.194 | 1.144 | 0.589 |
| (0.1, 0.9, mid) | 24.9 | 5.8 | 9.4 | 9.7 | 12.4 | 0.160 | 0.199 | 1.147 | 0.510 |
| (0.1, 0.5, all) | 27.2 | 5.8 | 8.4 | 9.7 | 12.8 | 0.157 | 0.182 | 1.154 | 0.605 |
| (0.1, 0.5, low) | 25.8 | 5.0 | 9.7 | 9.4 | 12.5 | 0.188 | 0.182 | 1.180 | 0.640 |
| (0.1, 0.5, high) | 25.6 | 4.6 | 6.1 | 8.9 | 11.3 | 0.147 | 0.197 | 1.121 | 0.573 |
| (0.5, 0.5, mid) | 22.5 | 4.5 | 5.1 | 8.6 | 10.2 | 0.393 | 0.235 | 1.170 | 0.636 |
| (0.9, 0.5, mid) | 23.4 | 5.0 | 5.2 | 8.0 | 10.4 | 0.534 | 0.290 | 1.091 | 0.597 |
Table 6. DISP parameter sweep on Pascal VOC. Values are in percent, followed by the four representation descriptors. The sweep is exploratory and includes target-domain scores.
| VOC | Clipart | Water. | COCO | Mean | |||||
| 0.1 | 26.8 | 5.4 | 7.8 | 9.7 | 12.4 | 0.134 | 0.201 | 1.111 | 0.517 |
| 0.3 | 27.8 | 6.1 | 9.8 | 9.7 | 13.3 | 0.165 | 0.227 | 1.128 | 0.551 |
| 0.6 | 24.8 | 6.6 | 9.1 | 8.7 | 12.3 | 0.233 | 0.280 | 1.153 | 0.578 |
| 0.9 | 22.9 | 4.5 | 8.0 | 8.0 | 10.9 | 0.326 | 0.368 | 1.160 | 0.585 |
Table 7. AmpMix sweep on Pascal VOC. Values are in percent, followed by the four representation descriptors.
The DISP sweep illustrates a clear magnitude effect in the current runs: increasing from 0.1 to 0.9 raises from 0.158 to 0.534 in the mid-band setting, while mean OOD decreases from 13.7% to 10.4%. Changes in at fixed alter the descriptor profile without a simple monotonic trend in . Frequency selection also changes the profile, supporting the use of as an intervention variable. These results should be read as sensitivity observations, not as evidence that any one direction in descriptor space is intrinsically optimal.
Computational Analysis
DISP adds no trainable parameters and is bypassed during inference. The reported model size remains approximately 2.51M parameters. The inference measurements below reflect the disabled-perturbation evaluation path and therefore should not be interpreted as training-time Fourier overhead.
| Method | Params (M) | Avg. Speed (ms) |
| Baseline | 2.51 | 2.46 0.21 |
| AmpMix | 2.51 | 2.49 0.20 |
| EFDM | 2.51 | 2.36 0.14 |
| FIESTA | 2.51 | 2.32 0.16 |
| DSU | 2.51 | 2.40 0.18 |
| MixStyle | 2.51 | 2.38 0.14 |
| NPOriginal | 2.51 | 2.34 0.15 |
| CCFP | 2.51 | 2.29 0.15 |
| UniFreqSDG | 2.52 | 2.28 0.13 |
| DISP | 2.51 | 2.66 0.18 |
Table 8. Model size and representative inference measurements.
Qualitative Feature-Map Analysis
The feature-map visualization complements the descriptor measurements. EFDM is a clear geometric outlier on VOC, with and . The corresponding shallow-layer activations show more substantial changes in localized response patterns than methods occupying the central part of the descriptor space. We treat this only as a qualitative illustration of the measured displacement, not as direct evidence of semantic distortion or causality.
DISP remains within the moderate-displacement range observed among the compared competitive configurations. Its representative feature responses change while retaining recognizable object and boundary structure in the displayed examples.

Figure 8. Representative shallow-layer feature-map responses. EFDM shows substantially larger measured displacement, while DISP exhibits more moderate changes consistent with its descriptor profile.
References
- [1]Rita Aitelhaj, Badr-Eddine Benelmostafa, and Hicham Medromi. Weathering the storm: Evaluating the generalization of recent object detection models for UAV power line inspection of insulators in challenging meteorological conditions. In New Technologies, Artificial Intelligence and Smart Data, pp. 392–408. Springer Nature Switzerland. ISBN 978-3-032-14964-0. doi: 10.1007/978-3-032-14964-0_38.
- [2]Badr-Eddine Benelmostafa, Rita Aitelhaj, and Hicham Medromi. YOLOv8-eRFD-AP: A novel domain generalization model for UAV-based insulator inspection under adverse weather conditions. 13:135336–135358. ISSN 2169-3536. doi: 10.1109/ACCESS.2025.3593201. URL https://ieeexplore.ieee.org/document/11097281.DOI
- [3]Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M. Saquib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. pp. 17732–17742. URL https://openaccess.thecvf.com/content/CVPR2024/html/Danish_Improving_Single_Domain-Generalized_Object_Detection_A_Focus_on_Diversification_and_CVPR_2024_paper.html?utm_source=chatgpt.com.arxiv.org/abs/2405.14497
- [4]Qi Fan, Mattia Segu, Bernt Schiele, Dengxin Dai, Yu-Wing Tai, and Chi-Keung Tang. Robust object detection with domain-invariant training and continual test-time adaptation. 133(10):6768– 6793. a. ISSN 1573-1405. doi: 10.1007/s11263-025-02465-9. URL https://doi.org/10.1007/s11263-025-02465-9.
- [5]Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Normalization perturbation: A simple domain generalization method for real-world domain shifts, b. URL http://arxiv.org/abs/2211.04393.
- [6]Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. URL http://arxiv.org/abs/2007.01434.
- [7]Zhenwei He, Hongsu Ni, Ling Lu, and Luxin Jiang. Single-domain generalized object detection by balancing domain diversity and invariance. 180:114184. ISSN 0031-3203. doi: 10.1016/j.patcog.2026.114184. URL https://www.sciencedirect.com/science/article/pii/S0031320326011490.
- [8]Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. URL http://arxiv.org/abs/1803.11365. version: 1.
- [9]Chenming Li, Daoan Zhang, Wenjian Huang, and Jianguo Zhang. Cross contrasting feature perturbation for domain generalization, a. URL http://arxiv.org/abs/2307.12502.
- [10]Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Lingyu Duan. Uncertainty modeling for out-of-distribution generalization. b. URL https://openreview.net/forum?id=6HN7LHyZgGc.
- [11]Chuang Liu, Yichao Cao, Xiu Su, and Haogang Zhu. Universal frequency domain perturbation for single-source domain generalization. In Proceedings of the 32nd ACM International Conference on Multimedia, MM ’24, pp. 6250–6259. Association for Computing Machinery. ISBN 979-8-4007-0686-8. doi: 10.1145/3664647.3681536. URL https://dl.acm.org/doi/10.1145/3664647.3681536.
- [12]Kwanseok Oh, Eunjin Jeon, Da-Woon Heo, Yooseung Shin, and Heung-Il Suk. FIESTA: Fourier-based semantic augmentation with uncertainty guidance for enhanced domain generalizability in medical image segmentation. URL http://arxiv.org/abs/2406.14308.
- [13]A.V. Oppenheim and J.S. Lim. The importance of phase in signals. 69(5):529–541. ISSN 1558-2256. doi: 10.1109/PROC.1981.12022. URL https://ieeexplore.ieee.org/document/1456290.
- [14]Fengchun Qiao and Xi Peng. Uncertainty-guided model generalization to unseen domains. URL http://arxiv.org/abs/2103.07531.
- [15]Zhiqiang Tang, Yunhe Gao, Yi Zhu, Zhi Zhang, Mu Li, and Dimitris Metaxas. CrossNorm and Self-Norm for generalization under distribution shifts. URL http://arxiv.org/abs/2102.02811.
- [16]Chengguang Wang, Zijun Zhang, and Zhidan Zhou. Domain feature perturbation for domain generalization. In Ulle Endriss, Francisco S. Melo, Kerstin Bach, Alberto Bugarín-Diz, José M. Alonso-Moral, Senén Barro, and Fredrik Heintz (eds.), ECAI 2024. SAGE Publications, a. doi: 10.3233/FAIA240782. URL https://journals.sagepub.com/action/showAbstract.
- [17]Xiangsheng Wang, Kangcheng Bin, Xiaojian Wang, Pengcheng Wan, Ting Hu, and Ping Zhong. Adversarial spectral perturbation for single-domain generalized object detection. 171:112264, b. ISSN 0031-3203. doi: 10.1016/j.patcog.2025.112264. URL https://www.sciencedirect.com/science/article/pii/S0031320325009252.
- [18]Qinwei Xu, Ruipeng Zhang, Ziqing Fan, Yanfeng Wang, Yi-Yan Wu, and Ya Zhang. Fourier-based augmentation with applications to domain generalization. 139:109474. ISSN 0031-3203. doi: 10.1016/j.patcog.2023.109474. URL https://www.sciencedirect.com/science/article/pii/S003132032301747.
- [19]Yabin Zhang, Bin Deng, Ruihuang Li, Kui Jia, and Lei Zhang. Adversarial style augmentation for domain generalization, a. URL http://arxiv.org/abs/2301.12643.