Generatively Inferential Co-Training for Unsupervised Domain Adaptation

Abstract

Deep Neural Networks (DNNs) have greatly boosted the performance on a wide range of computer vision and machine learning tasks. Despite such achievements, DNN is hungry for enormous high-quality (HQ) training data, which are expensive and time-consuming to collect. To tackle this challenge, domain adaptation (DA) could help learning a model by leveraging the knowledge of low-quality (LQ) data (i.e., source domain), while generalizing well on label-scarce HQ data (i.e., target domain). However, existing methods have two problems. First, they mainly focus on the high-level feature alignment while neglecting low-level mismatch. Second, there exists a class-conditional distribution shift even features being well aligned. To solve these problems, we propose a novel Generatively Inferential Co-Training (GICT) framework for Unsupervised Domain Adaptation (UDA). GICT is based on cross-domain feature generation and a specifically designed co-training strategy. Feature generation adapts the representation at low level by translating images across domains. Co-training is employed to bridge conditional distribution shift by assigning high-confident pseudo labels on target domain inferred from two distinct classifiers. Extensive experiments on multiple tasks including image classification and semantic segmentation demonstrate the effectiveness of GICT approach.