Sliced Wasserstein Discrepancy for Unsupervised Domain Adaptation (1903.04064v1)

Abstract: In this work, we connect two distinct concepts for unsupervised domain adaptation: feature distribution alignment between domains by utilizing the task-specific decision boundary and the Wasserstein metric. Our proposed sliced Wasserstein discrepancy (SWD) is designed to capture the natural notion of dissimilarity between the outputs of task-specific classifiers. It provides a geometrically meaningful guidance to detect target samples that are far from the support of the source and enables efficient distribution alignment in an end-to-end trainable fashion. In the experiments, we validate the effectiveness and genericness of our method on digit and sign recognition, image classification, semantic segmentation, and object detection.

• The paper introduces a novel discrepancy measure using the Wasserstein distance to align outputs between labeled source and unlabeled target domains.
• Its optimal transport-based approach outperforms methods like Maximum Classifier Discrepancy with improvements up to 2.8% in accuracy and 25% mAP in object detection.
• The method demonstrates versatility across tasks including digit recognition, image classification, semantic segmentation, and object detection while maintaining low complexity.

Sliced Wasserstein Discrepancy for Unsupervised Domain Adaptation

The paper introduces a novel approach to unsupervised domain adaptation by integrating the distribution alignment capabilities of task-specific decision boundaries with the geometrically meaningful Wasserstein metric. The focus is on addressing covariate shift between a labeled source domain and an unlabeled target domain, a common challenge in domain adaptation.

Key Contributions

1. Sliced Wasserstein Discrepancy (SWD): The authors propose a novel discrepancy measure leveraging the Wasserstein distance. This measure captures dissimilarities between the outputs of task-specific classifiers and facilitates efficient distribution alignment as part of an end-to-end trainable system.
2. Theoretical Foundation: The approach is grounded on optimal transport theory, utilizing the Wasserstein distance to provide a more substantial notion of probability distribution alignment compared to other measures such as MMD.
3. Versatility Across Tasks: The method is validated on diverse tasks, including digit and sign recognition, image classification, semantic segmentation, and object detection, showing its robust applicability across different domains.

Experimental Results

The experimental validation encompasses both standard and more challenging setups for domain adaptation. Notably, the SWD method surpasses existing techniques, including adversarial framework-based approaches, in several notable benchmarks:

• Digit and Sign Recognition: Achieves significant improvements, notably outperforming Maximum Classifier Discrepancy (MCD) with an average accuracy increase of 2.8% across four domain shifts.
• Image Classification on VisDA: Outperforms MCD and other methods with a mean accuracy improvement, demonstrating SWD’s capability in handling synthetic to real-world adaptations.
• Semantic Segmentation: Extensive evaluations on datasets like GTA5 to Cityscapes reveal the SWD method's superior performance, notably without relying on any specialty heuristic assumptions or priors.
• Object Detection: Implements SWD effectively, yielding a 25% relative improvement in mAP over MCD in challenging synthetic-to-real transitions.

Implications and Future Work

The proposed method's integration of Wasserstein discrepancy into domain adaptation frameworks can potentially redefine strategies in various challenging tasks without the need for additional network complexity or extensive architecture modifications. The theoretical underpinning suggests several avenues for future exploration, such as domain randomization, open set adaptation, and zero-shot domain adaptation.

The SWD method particularly shines in its ability to utilize the decision boundary strategically, allowing for more precise adaptation and generalization without the need for elaborate tasks. Future research can explore extending this framework further by integrating it with other unsupervised and semi-supervised learning paradigms, potentially opening new pathways in artificial intelligence domain adaptation capabilities.

This paper provides a solid foundation for ongoing research and could act as a launchpad for innovations in integrating optimal transport with deep learning-based domain adaptation approaches.