PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking Domain Generalization Baselines

Francesco Cappio Borlino, Antonio D'Innocente, Tatiana Tommasi

arXiv:2101.09060v2cs.CVcs.LG

TL;DR

Domain generalization concerns models deployed on domains different from their training data, while comparisons between its methods and data augmentation have been limited. This paper studies style-transfer augmentation and its integration with existing approaches. Experiments indicate that the original advantage of those approaches almost always disappears relative to the augmented baseline, with a stated boundary when domain shift is not due to style differences.

  • Problem

    Domain generalization methods face deployment on unknown domains, and recent feature- and model-based approaches have often not been benchmarked against data augmentation.

    Source: §I · p. 1

  • Method

    The paper uses style-transfer data augmentation and evaluates four domain-generalization methods under original and stylized augmentation settings, with tailored integrations.

    Source: §I · p. 1, B. Comparison methods · p. 3

  • Results

    Across PACS and OfficeHome, style augmentation improves the considered methods, but their original advantage over the augmented baseline almost always disappears; on VLCS, style transfer does not support generalization when shift is not due to style.

    Source: §I · p. 1, D. Results analysis · pp. 4–5

  • Takeaways & Limitations

    The authors call for domain-generalization strategies that take advantage of variability introduced by cross-domain style transfer.

    Source: §V · p. 6

    On VLCS, where the domain shift is not attributed to style differences, style-transfer augmentation does not support generalization.

    Source: D. Results analysis · p. 5

Abstract

from arXiv · show

Despite being very powerful in standard learning settings, deep learning models can be extremely brittle when deployed in scenarios different from those on which they were trained. Domain generalization methods investigate this problem and data augmentation strategies have shown to be helpful tools to increase data variability, supporting model robustness across domains. In our work we focus on style transfer data augmentation and we present how it can be implemented with a simple and inexpensive strategy to improve generalization. Moreover, we analyze the behavior of current state of the art domain generalization methods when integrated with this augmentation solution: our thorough experimental evaluation shows that their original effect almost always disappears with respect to the augmented baseline. This issue open new scenarios for domain generalization research, highlighting the need of novel methods properly able to take advantage of the introduced data variability.

I. INTRODUCTION

Domain generalization addresses deployment to unknown domains, but prior work often avoids comparing feature- and model-based methods with data augmentation. This paper presents style-transfer augmentation and tests how it interacts with existing approaches.

I. INTRODUCTION

  • Domain generalization assumes target data are available only at deployment, unlike domain adaptation, which uses unlabeled target data during training,,.Existing strategies seek invariances at the feature or model level,,.Source: §I · p. 1
  • Recent feature- and model-based methods often omit comparisons with data augmentation, motivating the paper's effort to clarify their relative performance,, [9],.The authors describe those approaches as treating augmentation as an unfair competitor because it extends the training set.Source: §I · p. 1
  • AdaIN-based style transfer augments source data while preserving semantic content, and the authors report top target results against existing approaches.The approach combines semantic and texture information from available source data.Source: §I · p. 1
  • Integrated domain-generalization methods almost always lose their original advantage over the style-augmented baseline.The authors also report that those methods benefit in absolute performance from source augmentation.Source: §I · p. 1

II. RELATED WORK

  • Feature-alignment methods reduce distances among source domains using discrepancy constraints, metric learning, or adversarial domain classifiers,,.Source: §II · p. 1
  • Related work includes meta-learning across meta-train and meta-test sources,,, self-supervised auxiliary tasks,, and augmentation from color changes to style transfer,,,,,,.Style-transfer approaches may use GAN architectures or data statistics.Source: §II · pp. 1–2

III. SOURCE AUGMENTATION BY STYLE TRANSFER

The paper studies multi-source domain generalization, where labeled source data are used to generalize to an unknown target. It augments source images by transferring style while retaining their labels and semantic content.

III. SOURCE AUGMENTATION BY STYLE TRANSFER

  • The setting uses multiple available sources and seeks generalization to an unknown target domain.Source: §III · p. 2
  • A style-transfer model maps each source image to a stylized version that retains its label, enriching training data while preserving semantic content.The paper describes a two-step process: learn the transfer model on source data, then use it for augmentation during classifier training.Source: §III · p. 2

A. Training the Style Transfer Model

The style-transfer model uses AdaIN to combine content and style features, then decodes the result into an image. Content and style losses guide reconstruction, with separate controls for training loss weight and test-time stylization.

A. Training the Style Transfer Model

  • AdaIN re-normalizes content features to match the style features' channel-wise means and standard deviations, then decodes the resulting representation.The encoder extracts features from the content and style images, and a decoder maps the combined features back to image space.Source: A. Training the Style Transfer Model · p. 2
  • The content loss measures the final feature output, while the style loss measures differences in mean and standard deviation across encoder-layer activations.Both compare features re-extracted from the stylized output with the target representation.Source: A. Training the Style Transfer Model · p. 3
  • The style-loss weight is generally fixed at λ = 10, while α controls the content-style trade-off at test time from content reconstruction (α = 0) to maximum stylization (α = 1).Source: A. Training the Style Transfer Model · p. 3

B. Style Transfer as Data Augmentation

During classifier training, batches draw samples across source domains, and each sample may be stylized using another batch instance as the style provider. A probability controls whether the stylized image replaces the original.

B. Style Transfer as Data Augmentation

  • Each batch sample can serve as content while another instance in the same batch is randomly selected as style provider, including from the same or a different source domain.The stylized image replaces the original with probability p.Source: B. Style Transfer as Data Augmentation · p. 3

IV. EXPERIMENTS

The experiments assess style-transfer augmentation against standard domain-generalization methods and evaluate their combination. They use three benchmarks with distinct class and domain characteristics, training on three source domains and testing on the held-out domain.

IV. EXPERIMENTS

  • The evaluation examines whether style-transfer augmentation improves the baseline, how it compares with recent DG methods, and how well the approaches combine.The study uses three standard benchmark datasets that differ in class count and covered domains.Source: §IV · p. 3, A. Datasets · p. 3

A. Datasets

  • PACS has 7 classes across Photo, Art Painting, Cartoon, and Sketch; models train on three source domains and test on the left-out domain.Validation splits support model selection.Source: A. Datasets · p. 3
  • OfficeHome covers 65 classes across Art, Clipart, Product, and Real-World, using a random 90–10 train-validation split for source-domain training and a held-out target.Validation images are used for model selection.Source: A. Datasets · p. 3
  • VLCS combines four datasets with 5 object categories; its domains are real-world photos, with shifts including camera, illumination, and viewpoint differences.Training uses three source domains; the held-out domain’s predefined test split is used for testing, and results average three runs.Source: A. Datasets · p. 3

B. Comparison methods

The comparison evaluates four domain-generalization methods with standard and style-transfer augmentation, alongside the baseline and Mixup. Integration choices account for whether methods require source-domain separation.

B. Comparison methods

  • DG-MMLD needs no source-domain labels, so style transfer is applied to content images with probability p, as for the Baseline.Random images within a batch provide the styles.Source: B. Comparison methods · p. 3
  • Epi-FCR relies on source labels to pair network modules by domain, so style mixing is avoided in that stage and applied when all source data are combined.The modules are the feature extractor and classifier; each is paired with a badly tuned partner for the domain in the current episode.Source: B. Comparison methods · pp. 3–4
  • The classifier pipeline borrows style from other images; the Baseline uses all source data, with Original flips and crops and Stylized adding style transfer.Four DG methods are evaluated under both augmentation settings.Source: B. Comparison methods · p. 3, Figure 2 · p. 4
  • DDAIG’s label classifier trains on original and synthetic source data, further extended with style-transfer-augmented samples.Its transformation network produces samples that retain the original label while fooling a domain classifier.Source: B. Comparison methods · p. 4
  • Rotation adds rotation recognition as a self-supervised task; style augmentation is straightforward because training does not use domain labels.The rotation task predicts one of four angles: 0°, 90°, 180°, or 270°.Source: B. Comparison methods · p. 4
  • Mixup serves as a further reference, tested by interpolating examples at pixel and feature levels.Its interpolation-strength parameter γ recovers the Baseline at γ = 0.Source: B. Comparison methods · p. 4

C. Training setup

The training setup uses AdaIN for style transfer and AlexNet or ResNet18 for classification. The authors also implement Rotation and run or reuse the other methods’ code and results.

C. Training setup

  • The style-transfer model A uses AdaIN with a VGG backbone and is trained for 20 epochs at a learning rate of 5e-5.The experiment-specific α and p values are reported in the result-table captions; sensitivity is analyzed in Section IV-E.Source: C. Training setup · p. 4
  • The classification model uses AlexNet and ResNet18; Baseline, Rotation, and Mixup train for 30k iterations with SGD and 96-image batches.Each batch contains 32 images per source domain; the learning rate is 0.001, weight decay is 0.0001, η is 0.5, and γ is 0.4.Source: C. Training setup · p. 4
  • Rotation is implemented by adding a rotation-recognition branch; for DG-MMLD, Epi-FCR, and DDAIG, the authors use the original code and report published results when possible.Results obtained by running the authors’ code are marked with an asterisk.Source: C. Training setup · p. 4

D. Results analysis

Across PACS and OfficeHome, style-transfer augmentation strengthens the baseline enough that state-of-the-art domain-generalization methods no longer outperform it. VLCS marks a boundary: style-transfer augmentation does not support generalization when domain shift is not due to style differences.

PACS

  • More than 5 percentage points: on PACS, stylized source data improves Baseline performance, while the considered DG methods cannot outperform the augmented Baseline.This pattern is reported for AlexNet and ResNet18; the improvement is higher for Art Painting, Cartoon, and Sketch than Photo.Source: D. Results analysis · p. 4, Table I · p. 5

OfficeHome

  • On OfficeHome, the Stylized Baseline obtains the best accuracy, outperforming competing state-of-the-art methods even when they also use source augmentation.The improvement from style-transfer augmentation is more limited here than on PACS.Source: D. Results analysis · pp. 4–5, Table II · p. 5

VLCS

  • On VLCS, style-transfer source augmentation does not support generalization because this testbed’s domain shift is not originally due to style differences.The reported results use AlexNet.Source: Table III · p. 5, D. Results analysis · p. 5

Mixup

  • Across the considered datasets, Mixup does not generalize across domains and may perform worse than the Original Baseline; its results remain below DG methods with or without style-based augmentation.Of the pixel and feature variants, only feature Mixup shows some advantage on PACS, so the authors use it in the other tests.Source: D. Results analysis · p. 5

E. Analysis of AdaIN hyperparameters

The PACS AlexNet analysis finds that strong style transfer is important for adding appearance variability, with the best results generally at α = 1. The section also illustrates style transfer across source-trained AdaIN variants and compares AdaIN training strategies.

Hyperparameter effects

  • With α = 1, PACS AlexNet results are generally best regardless of p; low α makes style transfer too weak to change appearance effectively and add variability.Figures 3 and 4 vary α and p in turn while holding the other fixed.Source: E. Analysis of AdaIN hyperparameters · p. 5
  • When α is high enough, even a small p can perform well, with the best results at p = 0.5 or p = 0.75.This finding concerns PACS AlexNet.Source: E. Analysis of AdaIN hyperparameters · p. 5

AdaIN examples

  • The example contrasts a pretrained MS-COCO/WikiArt AdaIN translation with source-trained AdaIN translations using Art Painting, Cartoon, Sketch, and Photo as style sources.The content image is from PACS Photo and the style image is from PACS Art Painting.Source: Figure 5 · p. 6

AdaIN training strategies

  • Table IV compares AdaIN training strategies.The supplied table caption does not state the comparison’s outcome.Source: Table IV · p. 6

F. Style transfer from external data vs source data

This section compares source-trained AdaIN variants with an AdaIN model pretrained on MSCOCO and WikiArt. On PACS AlexNet, augmentation with the pretrained model has only a slight, statistically nonsignificant accuracy edge over the Stylized Baseline.

F. Style transfer from external data vs source data

V. CONCLUSIONS

The paper evaluates a simple style-transfer augmentation strategy and tailored integrations with existing domain-generalization methods. The considered methods improve over their unaugmented versions but show no improvement over the new augmented baseline, motivating strategies that use the introduced variability.

V. CONCLUSIONS

  • Although the considered domain-generalization methods improve over their unaugmented versions, they show no improvement over the new data-augmented baseline.Source: §V · p. 6
  • The paper reports that its simple style-transfer augmentation overcomes existing competitors and uses tailored strategies to combine the augmentation with relevant domain-generalization methods.Source: §V · p. 6
  • The authors call for novel strategies able to take advantage of data variability introduced by cross-domain style transfer.Source: §V · p. 6

Related papers

Loading 2101.09060v2…