Domain Adaptation via Image Style Transfer

Abstract

While recent growth in modern machine learning techniques has led to remarkable strides in computer vision applications, one of the most significant challenges facing learning-based vision systems is the scarcity of large, high-fidelity datasets required for training large-scale models. This has necessitated the creation of transfer learning and domain adaptation as a highly-active area of research, wherein the objective is to adapt a model trained on one set of data from a specific domain to perform well on previously-unseen data from a different domain. In this chapter, we use monocular depth estimation as a means of demonstrating a new perspective on domain adaptation. Most monocular depth estimation approaches either rely on large quantities of ground truth depth data, which is extremely expensive and difficult to obtain, or alternatively predict disparity as an intermediary step using a secondary supervisory signal leading to blurring and other artefacts. Training a depth estimation model using pixel-perfect synthetic depth images can resolve most of these issues but introduces the problem of domain shift from synthetic to real-world data. Here, we take advantage of recent advances in image style transfer and its connection with domain adaptation to predict depth from a single colour image based on training over a large corpus of synthetic data obtained from a virtual environment. Experimental results point to the impressive capabilities of style transfer used as a means of adapting the model to unseen data from a different domain.