Paper deep dive
Evaluation of Randomization through Style Transfer for Enhanced Domain Generalization
Dustin Eisenhardt, Timothy Schaumlöffel, Alperen Kantarci, Gemma Roig
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%
Last extracted: 4/10/2026, 3:17:24 AM
Summary
This paper presents a systematic empirical study on style transfer as a data augmentation strategy for domain generalization in computer vision. The authors resolve contradictions in prior literature regarding style pool diversity, texture complexity, and style source. They introduce 'StyleMixDG', a lightweight, model-agnostic augmentation recipe that leverages a large, diverse pool of artistic styles, achieving competitive performance on the GTAV to {BDD100k, Cityscapes, Mapillary Vistas} benchmark without architectural modifications.
Entities (7)
Relation Signals (3)
StyleMixDG → evaluatedon → GTAV
confidence 100% · Evaluated on the GTAV -> {BDD100k, Cityscapes, Mapillary Vistas} benchmark
StyleMixDG → usesalgorithm → AdaIN
confidence 95% · As a method for style transfer, we use AdaIN
StyleMixDG → usesdataset → Painter by Numbers
confidence 95% · We choose the Painter By Numbers dataset as a source of artistic images
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Deep learning models for computer vision often suffer from poor generalization when deployed in real-world settings, especially when trained on synthetic data due to the well-known Sim2Real gap. Despite the growing popularity of style transfer as a data augmentation strategy for domain generalization, the literature contains unresolved contradictions regarding three key design axes: the diversity of the style pool, the role of texture complexity, and the choice of style source. We present a systematic empirical study that isolates and evaluates each of these factors for driving scene understanding, resolving inconsistencies in prior work. Our findings show that (i) expanding the style pool yields larger gains than repeated augmentation with few styles, (ii) texture complexity has no significant effect when the pool is sufficiently large, and (iii) diverse artistic styles outperform domain-aligned alternatives. Guided by these insights, we derive StyleMixDG (Style-Mixing for Domain Generalization), a lightweight, model-agnostic augmentation recipe that requires no architectural modifications or additional losses. Evaluated on the GTAV $\rightarrow$ {BDD100k, Cityscapes, Mapillary Vistas} benchmark, StyleMixDG demonstrates consistent improvements over strong baselines, confirming that the empirically identified design principles translate into practical gains. The code will be released on GitHub.
Tags
Links
- Source: https://arxiv.org/abs/2604.05616v1
- Canonical: https://arxiv.org/abs/2604.05616v1
Trouble viewing inline? Open PDF directly →
Full Text
38,446 characters extracted from source content.
Expand or collapse full text
Evaluation of Randomization through Style Transfer for Enhanced Domain Generalization Dustin Eisenhardt 1,2,3[0009−0007−6777−9038] , Timothy Schaumlöffel 3,4[0009−0007−4748−3037] , Alperen Kantarci 3[0000−0002−4080−5538] , and Gemma Roig 3,4[0000−0002−6439−8076] 1 German Cancer Research Center (DKFZ), Heidelberg, Germany 2 German Cancer Consortium (DKTK), partner site Frankfurt/Mainz, a partnership between DKFZ and UCT Frankfurt-Marburg, Frankfurt am Main, Germany 3 Goethe University Frankfurt, Department of Computer Science, Frankfurt am Main, Germany 4 The Hessian Center for Artificial Intelligence (hessian.AI), Darmstadt, Germany dustin.eisenhardt@dkfz-heidelberg.de Abstract. Deep learning models for computer vision often suffer from poor generalization when deployed in real-world settings, especially when trained on synthetic data due to the well-known Sim2Real gap. Despite the growing popularity of style transfer as a data augmentation strat- egy for domain generalization, the literature contains unresolved con- tradictions regarding three key design axes: the diversity of the style pool, the role of texture complexity, and the choice of style source. We present a systematic empirical study that isolates and evaluates each of these factors for driving scene understanding, resolving inconsistencies in prior work. Our findings show that (i) expanding the style pool yields larger gains than repeated augmentation with few styles, (i) texture complexity has no significant effect when the pool is sufficiently large, and (i) diverse artistic styles outperform domain-aligned alternatives. Guided by these insights, we derive StyleMixDG (Style-Mixing for Do- main Generalization), a lightweight, model-agnostic augmentation recipe that requires no architectural modifications or additional losses. Evalu- ated on the GTAV → BDD100k, Cityscapes, Mapillary Vistas bench- mark, StyleMixDG demonstrates consistent improvements over strong baselines, confirming that the empirically identified design principles translate into practical gains. The code will be released on GitHub. 1 Introduction Deep learning models for computer vision often suffer from poor generalization when deployed in real-world settings, especially when trained on synthetic data due to the Sim2Real gap [30]. This issue is particularly acute in autonomous driving, where collecting diverse, safety-critical datasets is costly or impractical. arXiv:2604.05616v1 [cs.CV] 7 Apr 2026 2Authors Suppressed Due to Excessive Length Dataset Pre-ProcessingTraining Style Dataset Synthetic Dataset N-times Augmented Dataset 20% 80% Weighted Random Choice Photo-metric Distortion Model Fig. 1: Overview of StyleMixDG. Each source image is stylized offline using N randomly selected styles, yielding an N-times augmented dataset. During train- ing, images are sampled with 80% probability from the augmented set and 20% from the original, followed by photometric distortion. Critically, we operate under the domain generalization assumption, where no real-world target data is available during training. This distinguishes our setting from domain adaptation approaches and reflects scenarios where data collection is infeasible due to cost, safety, or privacy constraints. To address this, one of the most widely used strategies is data augmentation, which aims to increase the diversity of the training set and improve robustness to distribution shifts. Within this domain, style transfer has emerged as a promising technique [22,17] to introduce controlled variation in the visual appearance of training samples. However, the literature contains unresolved contradictions that hamper the prin- cipled application of style transfer for domain generalization. Prior studies differ in the size of the style pool employed (e.g., 8 in [31] or 15 in [22]), in whether texture complexity of the style images matters [22,3], and in the choice of style source: artistic [22,24,31], source-domain [18,2], or real-world [18]. These dis- crepancies leave practitioners without clear guidance on how to configure style transfer augmentation. The primary contribution of this work is a systematic empirical study that isolates and evaluates each of these design axes under controlled conditions, re- solving conflicting findings in prior work. Concretely, we investigate: (1) whether expanding the style pool leads to better generalization than repeated augmenta- tion with a small set, (2) whether texture complexity of the style images affects domain robustness, and (3) whether the type of style source has a measurable influence on performance. Our study yields clear, actionable answers to all three questions. Evaluation of Randomization through Style Transfer3 Based on these empirical findings, we derive StyleMixDG (Style-Mixing for Domain Generalization), a simple augmentation recipe that instantiates the de- sign principles identified by our study. StyleMixDG is lightweight and model- agnostic: it does not require changes to the model architecture or training pipeline, nor does it introduce additional loss functions such as the alignment losses used in [22,29,17]. Instead, it increases data variability by mixing source-domain im- ages with artistically stylized variants drawn from a large pool, augmenting each image multiple times, and applying online photometric distortions during train- ing (see Fig. 1). StyleMixDG thus serves as both a practical baseline and a validation that the empirically identified principles translate into competitive performance. In summary, our contributions are: (1) a systematic empirical analysis of style diversity, texture complexity, and style source in style transfer augmentation, re- solving inconsistencies in prior work; (2) actionable design principles for practi- tioners configuring style transfer for domain generalization; and (3) StyleMixDG, a simple augmentation recipe derived from these principles that achieves compet- itive results on the GTAV → BDD100k, Cityscapes, Mapillary Vistas bench- mark without architectural or training pipeline modifications. 2 Related Work Style transfer, the process of applying the visual appearance of one image to an- other while preserving content [15], has been widely adopted for data augmenta- tion [31,27,12]. In the context of domain generalization (DG), several works em- ploy style transfer to reduce overfitting to domain-specific statistics [29,17,24,22]. Yue et al. [29] and Lee et al. [17] use external real-world styles, while Peng et al. [22] and Somavarapu et al. [24] rely on artistic styles to maximize visual variance. Kim et al. [16] explore weakly-labeled web data as both training input and style source, and Li et al. [18] investigate source-domain and real-world texture styles. However, prior works use small style pools (e.g., 8 in [31], 15 in [22]) and differ in their treatment of texture complexity and style source, leaving key design choices unresolved. We build on this line of work by systematically analyzing these factors and proposing a lightweight augmentation framework that requires no architectural or loss modifications. 3 Method We investigate the impact of diversity through style transfer in domain general- ization. To this end, we build on a simple style transfer augmentation method presented in Sec. 3.1. We extend it by increasing diversity in multiple ways as described in Sec. 3.2 3.1 Style Transfer Augmentation We employ style transfer as an offline data augmentation strategy, which we refer to as Style Transfer Augmentation (STA), following [22,24]. STA stylizes 4Authors Suppressed Due to Excessive Length Table 1: Examples of Style Transfer Augmentation (STA). This table shows examples of applying STA to synthetic images using style images of varying complexity. The first row depicts the style images taken from the Painter by Numbers[9] dataset. Content ImageLow ComplexityMedium Complexity High Complexity each input image using a randomly selected style image. This approach does not require any modifications to the model architecture or training procedure. Instead, models are trained on input images sampled from the style-augmented dataset. We choose the Painter By Numbers dataset [9] as a source of artistic images. This dataset comprises over 100.000 paintings from many artists and therefore provides a variability in style images. Using artistic styles for randomization should introduce irregular and unrealistic textures into the training process. As a consequence, the network should learn more domain-invariant cues such as shape and spatial layout, as it becomes harder to rely on texture. As a method for style transfer, we use AdaIN [14], which performs style transfer by aligning channel-wise mean and variance of content features to those of the style image. The stylized features are then decoded back to pixel space. Some examples of source, style, and stylized images are shown in Tab. 1. 3.2 StyleMixDG Building on the style transfer augmentation baseline described above, we intro- duce four extensions motivated by the empirical analysis in Sec. 5, which we preview here for clarity. First, we sample from a substantially larger set of styles (10,000) than commonly used in prior work (e.g., 8 in [31] or 15 in [22]). Second, each image is augmented with multiple different style images rather than only one. Third, we apply online photometric distortion during training. Fourth, we mix style-augmented and original source-domain images at an 80:20 ratio. The mixing ratio and number of stylized variants per image are hyperparameters se- Evaluation of Randomization through Style Transfer5 Table 2: State-of-the-Art Comparison. Shows the mIoU achieved using different networks on the three benchmark datasets (BDD, CS, MV) and their average. Source-only refers to standard training on synthetic data without any domain generalization technique applied. Methods above the gray divider rely solely on augmentation without modifying the model architecture or training pipeline; methods below require substantial architectural or loss-level changes. Results are either taken from [21] as denoted by + , from [20] as denoted by → , and otherwise from the original paper. VariantResNet50ResNet101 BDD CSMV AvgBDD CSMV Avg Source-only (Ours) 17.20 23.51 33.62 24.78 24.83 28.70 37.92 30.48 StyleMixDG (Ours) 35.61 40.75 42.45 39.54 36.34 42.74 45.15 41.41 GTR [22]33.75 + 37.53 + 34.52 + 35.16 + 39.60 43.70 39.10 40.80 SAN-SAW [21]37.34 39.75 41.86 39.65 41.20 ∗ 45.30 ∗ 40.80 ∗ 42.40 ∗ WildNet [17]38.42 + 44.62 + 46.09 + 43.04 + 41.70 ∗ 45.80 ∗ 47.10 ∗ 44.90 ∗ DIDEX [20]40.90 ∗ 52.40 ∗ 49.20 ∗ 47.50 ∗ lected in preliminary experiments. Each of these choices is empirically validated in Sec. 5; their combination defines StyleMixDG and is summarized in Fig. 1. 4 Experimental Setup Having introduced our method and its key design elements, we investigate whether our lightweight style transfer method can be competitive to state-of-the-art do- main generalization methods. In the following, we describe the benchmark and training settings used. 4.1 Benchmark The GTAV→ BDD, CS, MV benchmark is a common proxy for assessment of domain generalization performance in semantic segmentation tasks[20,21,29,16,22,17]. In this benchmark, neural networks are trained on the synthetic GTAV [23] dataset for automotive semantic segmentation. The datasets Berkeley Deep Drive 100k (BDD) [28], Cityscapes (CS) [7] and Mapillary Vistas (MV) [19] serve as real-world counterparts to GTAV, on which the networks will be evaluated. Since these datasets’ label spaces are not completely identical, they are mapped to a set of 19 common classes (see appendix Tab. 6). The benchmark performance is summarized using the average of the mIoU [10] achieved by the same model on the three benchmark datasets. We compare our method against a baseline of training on synthetic data without applying any domain generalization techniques, which we refer to as source-only training in the following. 6Authors Suppressed Due to Excessive Length ImageGroundtruthSource-onlyStyleMixDG unlabeledroadsidewalkbuildingwallfencepoletraffic lighttraffic signvegetationterrainskypersonridercartruckbustrainmotorcyclebicycle Fig. 2: Qualitative Results. This figure shows qualitative results of applying the R101 Source-only and StyleMixDG variants to images from the target datasets. 4.2 Training Settings We employ ResNets [11] as backbones and DeepLabV2 [4] as semantic segmen- tation head as it is common for this benchmark [22,20]. The backbones are initialized using ImageNet [8] weights. We use stochastic gradient descent as optimizer with a batch size of 8, momentum of 0.9 and weight decay of 5e −4 . The learning rate is initially set to 5e −3 and scheduled using a polynomial learning rate scheduler [5] to 5e −5 at the end of training. The networks are trained for 80.000 iterations. We use a weighted cross-entropy loss where class weights are computed per sample as one minus the pixel frequency of each class. We pre-process the GTAV dataset by cropping each image into two patches of 1052x1052 pixels with minimal overlap and resizing them to 640x640 pixels. In terms of augmentations, we employ random mirror and Gaussian blur with a probability of 50%. For the latter, the radius is randomly chosen from the interval (0, 1). 5 Experiments In this section, we first benchmark our approach against state-of-the-art meth- ods (Sec. 5.1), then conduct ablation experiments to validate our design choices and analyze how various aspects of style transfer influence domain generaliza- tion. We start by comparing the effectiveness of the style transfer augmentation to standard augmentations in Sec. 5.2. Following experiments investigate how increased variation contributes to generalization (Sec. 5.3), and highlight the role of complexity and type of styles (Sec. 5.4 and Sec. 5.5). Sec. 5.6 concludes the style transfer experiments by verifying the contribution of each component, followed by demonstrating transferability across architectures (Sec. 5.7). Evaluation of Randomization through Style Transfer7 5.1 Results Before presenting the detailed empirical analysis that motivates our design choices (Sec. 5.2 to Sec. 5.6), we first show that the configuration they converge on, StyleMixDG, is competitive with the state of the art, establishing that the find- ings have practical relevance. We compare StyleMixDG against state-of-the-art methods in Tab. 2 5 . With ResNet-50, StyleMixDG achieves an average mIoU of 39.54, significantly outper- forming GTR [22] (35.16), which is the method most comparable to ours. This is notable because GTR requires training on three differently augmented images simultaneously with a consistency loss, substantially increasing computational cost, whereas StyleMixDG relies solely on augmentation through simple random sampling. With ResNet-101, StyleMixDG remains competitive (41.41 vs. GTR’s 40.80 average mIoU), though the per-dataset picture is mixed: StyleMixDG leads on Cityscapes and Mapillary Vistas, while GTR achieves a higher score on BDD100K. The narrowing gap suggests that gains from data-driven aug- mentation alone may saturate with increasing model capacity. Methods such as SAN-SAW [21], WildNet [17] and DIDEX [20] achieve higher scores but rely on substantially more complex mechanisms including alignment modules, addi- tional losses, and modified training pipelines. In contrast, StyleMixDG is entirely model-agnostic, easy to integrate and comes at low computational cost. Quali- tative results are shown in Fig. 2. In the following sections, we iteratively build our method, starting with the style transfer augmentation baseline. 5.2 Baseline To ground subsequent improvements, we compare three ResNet-18 training vari- ants: source-only training, Style Transfer Augmentation (STA) with 15 styles following [22], and Photometric Distortion (PMD) [6]. PMD is a standard color- jitter augmentation applying random brightness, saturation, hue and contrast changes. Unless otherwise noted, all following sections use ResNet-18 and report mean ± standard deviation over three runs with different seeds and, for style transfer variants, differently augmented datasets. As shown in Fig. 3, STA im- proves over source-only training, consistent with prior work [22]. However, PMD performs competitively despite its simplicity, raising the question of whether the added cost of offline stylization can be justified. The following sections show how increasing stylistic diversity addresses this. 5.3 More Variation, More Robustness Standard STA is limited by applying stylization only once per image and draw- ing from a small style pool. We explore two extensions: (i) STA-MultAug, which 5 Per-class mIoUs for the ResNet-50 and ResNet-101 models can be found in appendix Tab. 8 and Tab. 9 respectively. 8Authors Suppressed Due to Excessive Length BDDCSMVAvg 20 30 40 mIoU SourceSTAPMDMultAugInfty Fig. 3: Results using more variation in style transfer. Source: training without style-transfer augmentation. STA is the baseline Style Transfer Augmentation. It is compared against Photo-metric Distortion (PMD), multiple style augmen- tations per image (MultAug) and sampling style images from a large pool of styles (Infty). Shows the mIoU achieved using different networks on the three benchmark datasets (BDD, CS, MV) and their average. generates three stylized versions per image from the same 15 styles and ran- domly samples one per iteration, and (i) STA-Infty, which samples from 10,000 styles while keeping one stylized version per image. As shown in Fig. 3, both variants outperform STA and PMD. STA-Infty yields the largest improvement, confirming that style pool diversity has a stronger effect than repeated augmen- tation of the same styles. Crucially, the clear advantage of STA-Infty over PMD demonstrates that style transfer provides benefits beyond what standard color augmentations can achieve, justifying its additional pre-processing cost. 5.4 Texture Complexity is Irrelevant Style images of varying texture complexity produce qualitatively different styl- izations (cf. Tab. 1): low-complexity styles create washed-out textures, while high-complexity styles may introduce artifacts. To test whether complexity af- fects generalization, we apply the Texture Complexity-based Painting Selection (TCPS) method from [22], which scores images by their ratio of small gradient pixels to total pixels. We partition the style dataset into low ([0, 0.5)), medium ([0.5, 0.75)) and high ([0.75, 1.0]) complexity subsets, sample 10,000 images from each, and train with STA. As shown in Tab. 3, all three variants achieve nearly identical performance (32.10, 32.17, 32.21 mIoU), indicating that texture com- plexity has no significant effect when sampling from a sufficiently large pool. This contrasts with the suggestion in [22] that complexity matters. We attribute the discrepancy to pool scale: at 10,000 styles, individual complexity effects av- erage out, making filtering unnecessary. We therefore exclude complexity-based selection from subsequent experiments. 5.5 The Choice of Style Dataset Matters While style transfer is widely used for domain generalization, prior works differ in their choice of style source: some rely on artistic styles [22,24,31], others on Evaluation of Randomization through Style Transfer9 Table 3: Results using different texture complexities. Compares sampling style images from low, medium and high-complexity styles with no complexity-based selection (Infty). The number of styles in the pool is identical. Shows the mIoU achieved using different networks on the three benchmark datasets (BDD, CS, MV) and their average. VariantBDDCSMVAvg Infty 28.99± 0.14 33.91± 0.24 34.45± 1.01 32.45± 0.28 Low28.11± 0.70 34.35± 0.25 33.83± 0.28 32.10± 0.17 Medium 28.46± 0.97 33.25± 0.83 34.81± 1.02 32.17± 0.76 High 29.22± 1.24 33.58± 1.64 33.84± 0.89 32.21± 0.94 BDDCSMVAvg 25 30 35 40 mIoU InftyIntra BDD*CS*MV*Mixed* Fig. 4: Results using different style sources. Style sources include the Painter- by-Numbers (Infty), GTAV (Intra) and benchmark datasets (BDD, CS, MV). Mixed denotes that all benchmark datasets are used to sample styles. Shows the mIoU achieved using different networks on the three benchmark datasets (BDD, CS, MV) and their average. The unsupervised domain adaptation settings are marked with an asterisk. source-domain styles [18,2] or real-world textures [18], yet the impact of this choice has not been systematically compared. We address this by evaluating artistic styles (STA-Infty) against source-domain stylization (STA-Intra, sam- pling styles from GTAV) and target-domain stylization (using BDD, CS, MV individually and combined). The latter constitutes unsupervised domain adap- tation (UDA), relaxing the DG setting for this analysis. As shown in Fig. 4 STA-Infty outperforms all domain-aligned variants. STA-Intra performs com- petitively but with higher variance. This suggests that a broad range of artistic styles, which introduce maximal distributional variation, is more effective than domain-aligned styles for generalization. 5.6 Ablation Study Having shown detailed investigations on the influence of style diversity, complex- ity and source, we provide an ablation of our method StyleMixDG. Tab. 4 shows that all building blocks contribute small improvements. Their combination leads to the strongest generalization performance observed in our experiments. 10Authors Suppressed Due to Excessive Length Table 4: Results using combinations of previous methods. Infty: increase the number of styles to sample from, Mixing: mix non-stylized and stylized im- ages, MultAug: use multiple styles per image, PMD: apply photo-metric distor- tion. Shows the mIoU achieved using different networks on the three benchmark datasets (BDD, CS, MV) and their average. The best and second-best entries are highlighted in bold and italics respectively. Infty Mixing MultAug PMD BDDCSMVAverage ✗ 17.07± 0.60 20.43± 0.38 31.87± 0.31 23.12± 0.32 ✓✗ 28.99± 0.14 33.91± 0.24 34.45± 1.01 32.45± 0.28 ✓✗✓ 29.22± 0.86 34.87± 0.89 34.58± 0.45 32.89± 0.35 ✓✗ 29.13± 0.84 33.57± 1.02 36.19± 0.66 32.97± 0.47 ✓✗✓✗ 29.09± 0.50 34.91± 0.13 34.77± 0.24 32.92± 0.20 ✓✗ 29 .81 ± 0 .12 36.03 ± 0 .39 35.66± 0.65 33 .83 ± 0 .30 ✓ 30.28± 0.68 37.12± 0.23 35 .85 ± 0 .76 34.41± 0.45 Table 5: Architecture Comparison. Source: non-stylized dataset, PMD: photo- metric distortion is applied, StyleMixDG: our method. Shows the mIoU achieved using different networks on the three benchmark datasets (BDD, CS, MV) and their average. The best entries are highlighted in bold. VariantResNet50 UperNetDeIT-S16 UperNet BDD CSMVAvgBDD CSMVAvg Source30.54 28.79 35.04 31.4633.74 33.04 38.63 35.14 PMD34.56 39.83 39.30 37.8932.87 38.69 37.80 36.45 StyleMixDG 38.23 41.07 42.75 40.68 39.52 38.19 42.29 40.00 5.7 Extension to Transformers To verify that our findings transfer beyond CNNs, we train an UperNet [26] seg- mentation head with both ResNet-50 and DeiT-S16 [25] backbones, following the training configuration of MMSegmentation [6]. DeiT-S16 is a vision transformer of comparable size to ResNet-50 (22.05M vs. 25.56M parameters). As shown in Tab. 5, StyleMixDG improves the DeiT-S16 baseline from 35.14 to 40.00 mIoU, confirming its effectiveness on transformer architectures. Interestingly, PMD slightly degrades DeiT-S16 performance on BDD100K and Mapillary Vis- tas while improving it on Cityscapes. This is consistent with findings that vision transformers exhibit reduced texture bias compared to CNNs [1], making color- targeting augmentations less beneficial or even counterproductive. StyleMixDG, which introduces broader distributional variation beyond color statistics, remains effective across both architecture families. Evaluation of Randomization through Style Transfer11 6 Limitations Style transfer can degrade small objects or render texture-distinguished classes indistinguishable, and rare classes such as train show limited improvement under StyleMixDG (cf. Tab. 8 and Tab. 9). Our evaluation is restricted to a single source domain (GTAV) and to semantic segmentation. Extending to additional synthetic sources, other perception tasks such as object detection, and mixed synthetic-real training regimes are natural directions for future work. 7 Conclusion We presented a systematic empirical study of style transfer augmentation for domain generalization in autonomous driving, addressing three open questions in the literature. Our controlled experiments show that (i) expanding the style pool has a larger effect than repeated augmentation with few styles, (i) tex- ture complexity filtering is unnecessary when the pool is sufficiently large, and (i) diverse artistic styles outperform domain-aligned alternatives. These find- ings resolve conflicting claims in prior work and provide actionable guidance for practitioners applying style transfer to domain generalization. As a practical instantiation of these principles, we derived StyleMixDG, a lightweight, model- agnostic augmentation recipe that achieves competitive results on the GTAV → BDD100k, Cityscapes, Mapillary Vistas benchmark without requiring archi- tectural or training pipeline modifications. This confirms that principled con- figuration of a simple augmentation strategy can rival more complex domain generalization methods. References 1. Angarano, S., Martini, M., Salvetti, F., Mazzia, V., Chiaberge, M.: Back-to-bones: Rediscovering the role of backbones in domain generalization. Pattern Recognition 156 (2024) 2. Borlino, F.C., D’Innocente, A., Tommasi, T.: Rethinking domain generalization baselines. In: ICPR (2021) 3. Cakir, D., Arica, N.: Style transfer to enhance data augmentation for facial action unit detection. In: ROSENET (2024) 4. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Se- mantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI (2018) 5. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint:1706.05587 (2017) 6. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020) 7. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016) 8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large- Scale Hierarchical Image Database. In: CVPR (2009) 12Authors Suppressed Due to Excessive Length 9. small yellow duck, Kan, W.: Painter by numbers. https://kaggle.com/ competitions/painter-by-numbers (2016), kaggle 10. Everingham, M., Eslami, S.M.A., Van Gool, L., Williams, C.K.I., Winn, J., Zisser- man, A.: The pascal visual object classes challenge: A retrospective. IJCV (2015) 11. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 12. Hong, M., Choi, J., Kim, G.: Stylemix: Separating content and style for enhanced data augmentation. In: CVPR (2021) 13. Hoyer, L., Dai, D., Van Gool, L.: Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In: CVPR (2022) 14. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: ICCV (2017) 15. Jing, Y., Yang, Y., Feng, Z., Ye, J., Yu, Y., Song, M.: Neural style transfer: A review. IEEE TVCG (2020) 16. Kim, N., Son, T., Pahk, J., Lan, C., Zeng, W., Kwak, S.: Wedge: web-image assisted domain generalization for semantic segmentation. In: ICRA (2023) 17. Lee, S., Seong, H., Lee, S., Kim, E.: Wildnet: Learning domain generalized semantic segmentation from the wild. In: CVPR (2022) 18. Li, Y., Zhang, D., Keuper, M., Khoreva, A.: Intra-source style augmentation for improved domain generalization. In: WACV (2023) 19. Neuhold, G., Ollmann, T., Rota Bulo, S., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: ICCV (2017) 20. Niemeijer, J., Schwonberg, M., Termöhlen, J.A., Schmidt, N.M., Fingscheidt, T.: Generalization by adaptation: Diffusion-based domain extension for domain- generalized semantic segmentation. In: WACV (2024) 21. Peng, D., Lei, Y., Hayat, M., Guo, Y., Li, W.: Semantic-aware domain generalized segmentation. In: CVPR (2022) 22. Peng, D., Lei, Y., Liu, L., Zhang, P., Liu, J.: Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE TIP (2021) 23. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: ECCV (2016) 24. Somavarapu, N., Ma, C.Y., Kira, Z.: Frustratingly simple domain generalization via image stylization. arXiv preprint:2006.11207 (2020) 25. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: ICML (2021) 26. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: ECCV (2018) 27. Yang, Q., Ji, C., Luo, H., Li, P., Ding, Z.: Data augmentation through random style replacement. arXiv preprint:2504.10563 (2025) 28. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In: CVPR (2020) 29. Yue, X., Zhang, Y., Zhao, S., Sangiovanni-Vincentelli, A., Keutzer, K., Gong, B.: Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In: ICCV (2019) 30. Zhao, W., Queralta, J.P., Westerlund, T.: Sim-to-real transfer in deep reinforce- ment learning for robotics: a survey. In: IEEE SSCI (2020) 31. Zheng, X., Chalasani, T., Ghosal, K., Lutz, S., Smolic, A.: Stada: Style transfer as data augmentation. In: VISIGRAPP (2019) Evaluation of Randomization through Style Transfer13 A Implementation Details A.1 Style Transfer Augmentation We use the style transfer implementation and pre-trained networks from https: //github.com/naoto0804/pytorch-AdaIN. We only select paintings that have a width and height of 512 pixels or more. Before encoding the style images, we take the largest square-center crop of the style image and resize it to 512x512 pixels. We extract the features of the content images at the original size, i.e. 1052x1052 pixels. As the decoded image has a resolution of 1056x1056 pixels, we resize it to match the original size. A.2 Texture Complexity-based Painting Selection Pixels are classified as smooth if their gradient representation is smaller than a threshold ε = 20. As the resolution of the paintings influences the area and therefore also the complexity, we use uniform painting resolutions for complexity computation. In particular, we use the largest square-center crop and resize it to a resolution of 512x512 pixels. After resizing, we convert the images to grayscale. We calculate the first discrete difference along the spatial axes to compute the pixel gradients: grad x (i,j) = I (i + 1,j)− I (i,j) grad y (i,j) = I (i,j + 1)− I (i,j) We apply zero padding to the size of 512 pixels squared. The gradients in the horizontal and vertical direction are combined by summing the squared gradi- ents: grad (i,j) = grad 2 x (i,j) + grad 2 y (i,j). Using this procedure, we reproduce the smooth versus unsmooth area separation shown by Peng et al. [22] in Tab. 7. It reproduces the smooth versus unsmooth area separation originally shown in GTR[22] to validate that our TCPS im- plementation yields consistent results. Since GTR does not provide sufficient implementation details, this replication serves as a verification step for our own implementation. A.3 Photo-metric Distortion Examples of Photo-metric Distortion are shown in Fig. 5 A.4 Model Architecture Even though the DeepLabV2[4] architecture is used in related works, there are subtle modifications to it. For example, it is common to use only the first two out of the four convolutional layers in the segmentation head (cf. [22,13], which we adopt. Other works freeze the first batch-normalization layer during training or 14Authors Suppressed Due to Excessive Length Table 6: Label Correspondences. Label correspondences used in our experiments. The Label column additionally shows the color for each class used in subsequent visualization. GTAV[23]BDD[28] CS[7]MV[19] roadroadroadbike lane, crosswalk - plain, lane marking - crosswalk, lane marking - general, man- hole, parking, road, service lane sidewalk sidewalk sidewalkcurb, curb cut, pedestrian area, sidewalk building building buildingbuilding wallwallwallwall fencefencefencefence polepolepole, polegrouppole, street light, utility pole traffic light traffic light traffic lighttraffic light traffic sign traffic sign traffic signtraffic sign (back), traf- fic sign (front), traffic sign frame vegetation vegetation vegetationvegetation terrainterrainterrainterrain skyskyskysky personpedestrian personperson riderriderriderbicyclist, motorcyclist, other rider carcarcarcar trucktrucktrucktruck Table continues on the next page. Evaluation of Randomization through Style Transfer15 Table 6: Label Correspondences. Label correspondences used in our experiments. The Label column additionally shows the color for each class used in subsequent visualization. GTAV[23]BDD[28] CS[7]MV[19] busbusbusbus traintraintrainon rails motorcycle motorcycle motorcyclemotorcycle bicyclebicyclebicyclebicycle unlabeled unlabeled bridge, caravan, dynamic, ego vehicle, ground, guard rail, license plate, out of roi, parking, rail track, rectifica- tion border, static, trailer, tunnel, unlabeled banner, barrier, bench, bike rack, billboard, bird, boat, bridge, car mount, caravan, catch basin, cctv camera, ego vehicle, fire hydrant, ground animal, guard rail, junc- tion box, mailbox, moun- tain, other vehicle, phone booth, pothole, rail track, sand, snow, trailer, trash can, tunnel, unlabeled, wa- ter, wheeled slow (a)(b)(c)(d) (e)(f)(g)(h) Fig. 5: Examples of Photo-Metric Distortion. (a) shows the source image that is augmented using photo-metric distortion in (b)-(h). 16Authors Suppressed Due to Excessive Length OriginalSmooth AreaUnsmooth Area 14405 (0.1537) 13473 (0.1411) 1425 (0.5252) 12582 (0.6923) 13170 (0.7896) 11151 (0.8427) Table 7: Smooth vs. Unsmooth Area Separation. Smooth and unsmooth pixels are filtered using the gradient representations. The images are accompanied by their id in the Painter by Numbers dataset [9] and their texture complexity. This figure is a validation of our implementation of TCPS as [22] does not provide sufficient implementation details. Evaluation of Randomization through Style Transfer17 use the 10-fold learning rate for the segmentation head (cf. [13]). We do not adopt these changes in order to keep the training setup as simple and reproducible as possible, and to avoid introducing confounding factors that could obscure the contribution of the augmentation strategy itself. 18Authors Suppressed Due to Excessive Length R50 SourceR50 StyleMixDG classbdd cs mv mean bdd cs mv mean +/- bicycle 01.59 08.58 19.82 10.00 17.29 20.53 24.82 20.88 +10.88 building 32.88 46.92 58.33 46.04 65.67 80.18 72.87 72.91 +26 .86 bus04.18 01.54 07.29 04.34 17.23 27.42 26.92 23.86 +19.52 car28.08 61 .92 72.13 54.04 73.10 78.36 77.23 76.23 +22.19 fence 12.94 14.06 21.26 16.09 20.03 16.96 24.80 20.60 +04.51 m. cycle 03.42 07.58 21.34 10.78 26.33 14.79 28.62 23.25 +12.47 person 25.43 40.49 49.42 38.45 40.85 59.66 55.88 52.13 +13.68 pole20.63 18.17 26.96 21.92 27.87 28.38 31.37 29.21 +07.29 rider03.35 07.41 17.71 09.49 17.11 22.37 30.70 23.39 +13.90 road35.84 32.46 65.37 44.56 78 .67 86.78 75.21 80.22 +35.66 sidewalk 10.81 13.11 26.24 16.72 33.46 42.19 42.65 39.43 +22.71 sky 69.42 52.16 76.46 66 .01 80.27 84 .26 81.94 82.16 +16.14 terrain 15.76 09.82 35.33 20.30 28.32 29.97 41.17 33.15 +12.85 t. light 25.74 23.38 36.55 28.56 31.11 29.51 35.25 31.96 +03.40 t. sign 19.51 07.25 23.42 16.73 27.29 21.56 35.12 27.99 +11.26 train00.00 00.03 10.52 03.52 00.00 02.64 05.68 02.77 −00.74 truck 07.07 08.35 25.60 13.67 16.34 17.94 26.89 20.39 +06.72 veget. 55.39 77.87 71 .55 68.27 68.58 82.85 76.02 75.82 +07.55 wall02.37 10.67 06.73 06.59 06.99 24.37 13.49 14.95 +08.36 mean 19.71 23.25 35.37 26.11 35.61 40.56 42.45 39.54 +13.43 Table 8: Detailed Evaluation of StyleMixDG on ResNet50. This table lists the per-class IoU achieved by the ResNet50 backbone using the source-only and StyleMixDG training settings. The best and second-best entries are highlighted in bold and italics respectively. Evaluation of Randomization through Style Transfer19 R101 SourceR101 StyleMixDG classbdd cs mv mean bdd cs mv mean +/- bicycle 09.32 13.69 29.96 17.65 23.11 22.29 28.45 24.62 +06.96 building 44.32 48.54 62.96 51.94 59.70 78.74 61.54 66.66 +14.72 bus02.89 12.69 08.46 08.01 09.93 25.67 22.96 19.52 +11.51 car50.18 64 .83 73.30 62.77 77.21 83.94 82.44 81.20 +18.43 fence 13.14 13.32 24.99 17.15 19.65 18.99 28.34 22.33 +05.18 m. cycle 14.05 12.86 28.12 18.34 23.24 18.38 35.32 25.65 +07.30 person 32.64 56.98 49.59 46.40 45.30 62.02 59.31 55.54 +09.14 pole22.44 18.91 30.93 24.09 32.59 34.62 36.61 34.61 +10.51 rider05.53 13.53 19.88 12.98 12.08 16.26 23.98 17.44 +04.46 road54.18 44.00 59.99 52.72 82.66 87.06 84.00 84.57 +31.85 sidewalk 19.09 23.23 27.88 23.40 35.88 46.80 46.24 42.97 +19 .57 sky 71.78 60.20 79.68 70 .55 77 .49 86.58 81.96 82.01 +11.46 terrain 16.98 17.99 33.23 22.73 32.29 33.82 40.04 35.38 +12.65 t. light 22.05 27.53 37.44 29.01 34.65 35.70 39.56 36.64 +07.63 t. sign 22.13 14.26 31.28 22.56 27.94 27.29 41.87 32.37 +09.81 train00.00 01.39 13.66 05.02 00.00 02.48 10.62 04.37 −00.65 truck 09.02 13.60 30.43 17.68 18.74 23.45 36.47 26.22 +08.54 veget. 60.56 80.00 73 .44 71.33 71.83 85 .05 76.49 77.79 +06.46 wall01.40 07.72 05.32 04.81 06.15 22.90 21.60 16.88 +12.07 mean 24.83 28.70 37.92 30.48 36.34 42.74 45.15 41.41 +10.93 Table 9: Detailed Evaluation of StyleMixDG on ResNet101. This table lists the per-class IoU achieved by the ResNet101 backbone using the source-only and StyleMixDG training settings. The best and second-best entries are highlighted in bold and italics respectively.