Reducing the Covariate Shift by Mirror Samples in Cross Domain Alignment

TL;DR

This paper introduces a novel mirror sample concept and mirror loss to improve cross-domain alignment in unsupervised domain adaptation, effectively reducing covariate shift without distorting distribution structures.

Contribution

It proposes the virtual mirror and mirror loss to better align distributions, addressing limitations of existing methods and providing theoretical guarantees.

Findings

Achieved superior performance on multiple benchmarks

Preserved the internal structure of distributions

Theoretically proven to reduce domain shift asymptotically

Abstract

Eliminating the covariate shift cross domains is one of the common methods to deal with the issue of domain shift in visual unsupervised domain adaptation. However, current alignment methods, especially the prototype based or sample-level based methods neglect the structural properties of the underlying distribution and even break the condition of covariate shift. To relieve the limitations and conflicts, we introduce a novel concept named (virtual) mirror, which represents the equivalent sample in another domain. The equivalent sample pairs, named mirror pairs reflect the natural correspondence of the empirical distributions. Then a mirror loss, which aligns the mirror pairs cross domains, is constructed to enhance the alignment of the domains. The proposed method does not distort the internal structure of the underlying distribution. We also provide theoretical proof that the mirror samples and mirror loss have better asymptotic properties in reducing the domain shift. By applying the virtual mirror and mirror loss to the generic unsupervised domain adaptation model, we achieved consistent superior performance on several mainstream benchmarks. Code is available at https://github.com/CTI-VISION/Mirror-Sample