Preface

While the proliferation of sensors being deployed in cell phones, vehicles, buildings, roadways, and computers allows for larger and more diverse information to be collected, the cost of acquiring labels for all these data remains extremely high. To overcome the burden of annotation, alternative solutions have been proposed in the literature to learn decision making models by exploiting unlabeled data from the same domain (data acquired in similar conditions as the targeted data) or also data from related but different domains (different datasets due to different conditions or provided by different customers). In many real-world machine learning scenarios, using only the data from the same domain might be insufficient and data or models borrowed from similar domains can significantly improve the learning process. Such a process, referred to as domain adaptation, aims to leverage labeled data in one or more related domains (sources), in order to build models for a target domain. Domain adaptation is particularly critical for service companies, where all machine learning components deployed in a given service solution should be customized for a new customer either by annotating new data or, preferably, by calibrating the models in order to achieve a contractual performance in the new environment. While adaptation across domains is a challenging task for many applications, in this book, we focus on solutions for visual applications. The aim of the book is to give a relatively broad view of the field by selecting a diverse set of methods which made different advances in the field. The book begins with a comprehensive survey of domain adaptation and transfer learning, including historical shallow methods, more recent methods using deep architectures, and methods addressing computer vision tasks beyond image categorization, such as detection, segmentation or visual attributes. Then, Chap. 2 gives a deeper look at dataset bias in existing datasets when different representations including features extracted from deep architectures are used. The rest of the book is divided into four main parts, following the same structure as the survey presented in Chap. 1. Part I is dedicated to shallow domain adaptation methods, beginning with the widely used Geodesic Flow Kernel (Chap. 3) and Subspace Alignment (Chap. 4). Both chapters propose solutions for selecting landmark samples in the source dataset. Chapter 5 presents domain-invariant embedding methods and Chap. 6