Visual Domain Adaptation via Feature Matching and Adaptive Classification

Abstract

In machine learning algorithms, it is often assumed that the training samples (source domain) and test samples (target domain) are collected in same conditions with same distributions. While in real-world problems, due to distribution shift across the source and target domains, machine learning algorithms can not build an optimal cross-domain model. In this way, domain adaptation methods are proposed to impove the performance of learning model with distribution difference. These methods, via reducing distribution shift across source and target data, build an adaptable model based on source domain where has maximum alignment with structure of target domain. Overall, to reduce distribution shift across domains, three types of solutions have been proposed. In first category, called feature matching, an adaptable subspaces for mapping data is built and in new subspaces, the distribution distance is minimized. In second category, called sample selection, the adaptable model is built on samples of source domain where have minimum distribution difference with target samples. In third category, called model matching, the classifier parameters created on source samples are adapted with structure of target domain. In this thesis, four groups of methods are proposed to reduce the distribution shift across domains. In first group, the source and target domains are mapped into new subspaces with maximum alignment of global, statistical and geometrical structures across domains. In second group, an adaptable classifier is built via geometrical and statistical distribution alignment, prediction error minimization and sample reweighting. In third group, a commn representation across domains is obtained via distribution difference reduction and class separability maximization. In the last group, at first the new representation of the source and target domains are built and then, the discriminant directions of the source and target domains are adapted by building an adaptable classifier in new representation. The proposed methods in this thesis were evaluated on benchmark visual datasets (in face, digit and object detection problems). The reported results highlight the considerable performance of the proposed approachs against other state-of-the-art domain adaptation methods.

Key Words : Unsupervised domain adaptation, Image classification, Feature matching,
Adaptive classifier, Subspace alignment, Distribution adaptation.