Source-linked AI summary
Best sources forward: domain generalization through source-specific nets
Massimiliano Mancini, Samuel Rota Bulò, Barbara Caputo, Elisa Ricci
TL;DR
Visual recognition systems need to generalize from source domains to unknown target domains, a setting less directly addressed by source–target Domain Adaptation. The paper introduces an end-to-end deep network with source-specific classifiers, target-conditioned fusion, and an optional domain-agnostic component; experiments on two benchmarks report state-of-the-art performance.
Problem
Domain Generalization seeks robust recognition across unknown target distributions, beyond methods focused on a specific source–target pair.
Method
The method uses source-specific classifiers, a domain-prediction branch to weight and fuse them for each input, and an optional domain-agnostic classifier.
Results
The proposed deep architecture outperforms state-of-the-art models on two benchmarks.
Takeaways & Limitations
Fusing multiple source models is an effective Domain Generalization strategy, with a domain-agnostic component improving accuracy in some settings.
Abstract
from arXiv · showhide
A long standing problem in visual object categorization is the ability of algorithms to generalize across different testing conditions. The problem has been formalized as a covariate shift among the probability distributions generating the training data (source) and the test data (target) and several domain adaptation methods have been proposed to address this issue. While these approaches have considered the single source-single target scenario, it is plausible to have multiple sources and require adaptation to any possible target domain. This last scenario, named Domain Generalization (DG), is the focus of our work. Differently from previous DG methods which learn domain invariant representations from source data, we design a deep network with multiple domain-specific classifiers, each associated to a source domain. At test time we estimate the probabilities that a target sample belongs to each source domain and exploit them to optimally fuse the classifiers predictions. To further improve the generalization ability of our model, we also introduced a domain agnostic component supporting the final classifier. Experiments on two public benchmarks demonstrate the power of our approach.
1. INTRODUCTION
Domain Generalization addresses visual recognition across unknown target domains, extending beyond methods designed for a specific source–target pair. The paper proposes an end-to-end deep architecture that combines source-specific classifiers using target-to-source similarities and optionally adds a domain-agnostic classifier.
- Visual recognition systems must generalize across domains because future target data may differ statistically from the source training data.
- Domain Adaptation typically addresses domain shift between one specific source and one specific target, whereas Domain Generalization targets arbitrary future domains.
- The proposed network replaces domain-invariant representations with multiple source-specific classifiers whose prediction scores are optimally fused for each target image.
- A domain-prediction branch estimates input similarity to source domains, assigns classifier weights, and merges their predictions in an end-to-end trainable architecture.
- The work belongs to the source-similarity approach to Domain Generalization but casts it into an end-to-end deep architecture that maintains source-specific representations.
2. DOMAIN GENERALIZATION WITH SOURCE-SPECIFIC CLASSIFIERS
The framework combines source-specific classifiers through weights predicted from each input, while sharing most network parameters across domain branches. It also adds domain-agnostic averaging and jointly trains semantic classification with domain prediction.
- Architecture: The model maps each input image to class predictions by combining outputs from N domain-specific classifiers.Each classifier is associated with one source domain.
- Architecture: A convolutional network implements parallel source-specific branches that share parameters except for the final classifier layer.This parameter sharing reduces the computational cost of maintaining separate domain models.
- Domain weighting: For source samples, the domain label sets the weight for the corresponding classifier to 1 and the remaining classifier weights to 0.This trains the source-specific classifiers using samples from their associated domains.
- Domain weighting: A domain prediction branch produces a normalized weight vector whose entries represent the input image’s probabilities of belonging to the source domains.The weights are constrained between 0 and 1 and sum to 1.
- Training: The architecture jointly minimizes semantic classification and domain prediction losses, with λ controlling their relative contribution.Both losses are implemented as cross-entropy losses.
- Domain-agnostic component: A domain-agnostic component mixes uniform classifier weighting with predicted domain weights to improve robustness when source-specific classifiers are weak.The trade-off is regulated by α, which randomly switches between the two weighting schemes during training.
3. EXPERIMENTS
Experiments on rotated-MNIST and PACS evaluate the proposed domain-generalization network across differing source–target conditions. The model outperforms prior methods, while source-specific classifier fusion and a domain-agnostic component show complementary effects.
- Datasets: The evaluation uses rotated-MNIST with one target rotation and the remaining rotations as sources, and PACS with three source domains and one target domain.PACS covers Photo, Art paintings, Cartoon, and Sketches, with substantial domain shift.
- Rotated-MNIST results: The proposed model outperforms all compared baselines on rotated-MNIST, with a particularly notable accuracy gain at 45°.The authors attribute the gain to reweighting source classifiers associated with nearby orientations.
- Rotated-MNIST results: The domain prediction branch assigns target samples mainly to source domains with the closest rotations, enabling predictions from appropriately combined source classifiers.Assignments are computed using the largest source-domain weight for each target sample.
- PACS results: On PACS, the proposed model outperforms previous methods, where domain shifts are more substantial and harder to capture than simple image perturbations.The comparison uses AlexNet-based deep models alongside prior methods using precomputed features.
- Sensitivity analysis: Across PACS sensitivity tests, α = 0 is usually strongest, while adding a domain-agnostic component with α = 0.25 improves accuracy in some settings such as Cartoon.The analysis compares source-specific classifiers with a domain-agnostic classifier and mixtures of the two.
4. CONCLUSIONS
The paper concludes that multiple domain-specific classifiers, combined by a domain prediction branch, provide an effective architecture for domain generalization. Experiments on two benchmarks show that the approach outperforms state-of-the-art models.
- Conclusion: The proposed architecture uses multiple domain-specific classifiers and selects their test-time combination according to input similarity to source domains.A domain-agnostic component is also included to improve performance.
- Conclusion: Experiments on two benchmarks demonstrate that the proposed deep architecture outperforms state-of-the-art models.