Geodesic flow kernel for unsupervised domain adaptation

TLDR

In visual recognition, variations in pose, illumination, or image quality create a mismatch between source and target domains, causing classifiers to perform poorly and motivating domain adaptation techniques that seek to correct this shift. This work introduces a kernel‑based method that exploits low‑dimensional structures inherent in vision data to address domain mismatch. The proposed geodesic flow kernel models domain shift by integrating infinitely many subspaces capturing geometric and statistical changes, automatically infers key parameters without cross‑validation, and employs a metric to assess and select the most adaptable source domain. Experiments on standard datasets demonstrate that this approach outperforms competing domain adaptation methods.

Abstract

In real-world applications of visual recognition, many factors - such as pose, illumination, or image quality - can cause a significant mismatch between the source domain on which classifiers are trained and the target domain to which those classifiers are applied. As such, the classifiers often perform poorly on the target domain. Domain adaptation techniques aim to correct the mismatch. Existing approaches have concentrated on learning feature representations that are invariant across domains, and they often do not directly exploit low-dimensional structures that are intrinsic to many vision datasets. In this paper, we propose a new kernel-based method that takes advantage of such structures. Our geodesic flow kernel models domain shift by integrating an infinite number of subspaces that characterize changes in geometric and statistical properties from the source to the target domain. Our approach is computationally advantageous, automatically inferring important algorithmic parameters without requiring extensive cross-validation or labeled data from either domain. We also introduce a metric that reliably measures the adaptability between a pair of source and target domains. For a given target domain and several source domains, the metric can be used to automatically select the optimal source domain to adapt and avoid less desirable ones. Empirical studies on standard datasets demonstrate the advantages of our approach over competing methods.