PROJECT TITLE :

A Review of Single-Source Deep Unsupervised Visual Domain Adaptation

ABSTRACT:

Deep neural networks have been shown to perform exceptionally well across a broad spectrum of benchmark vision tasks as a result of the use of large-scale labeled training datasets. In spite of this, it can be prohibitively expensive and time-consuming to acquire large quantities of labeled data in a number of different applications. Many people have tried to directly apply models that were trained on a large-scale labeled source domain to another domain that was either sparsely labeled or unlabeled in an effort to circumvent the problem of having insufficient labeled training data. Unfortunately, due to the presence of domain shift or dataset bias, the performance of direct transfer between domains is frequently subpar. The goal of the machine learning paradigm known as domain adaptation (DA) is to learn a model from one domain (the source domain) that can function effectively in another domain (the target domain), which is related to the source domain. This article provides a comprehensive review of the most recent single-source deep unsupervised DA methods that are focused on visual tasks, as well as a discussion of new research perspectives. In the first step of this process, we provide definitions of the various DA strategies and descriptions of the existing benchmark datasets. Discrepancy-based methods, adversarial discriminative methods, adversarial generative methods, and self-supervision-based methods are some of the different types of single-source unsupervised DA methods that we will now summarize and compare. In conclusion, we discuss future directions for research along with potential problems and their possible solutions.