AWADA: Foreground-focused adversarial learning for cross-domain object detection

Abstract

Object detection networks have achieved impressive results, but it can be challenging to replicate this success in practical applications due to a lack of relevant data specific to the task. Typically, additional data sources are used to support the training process. However, the domain gaps between these data sources present a challenge. Adversarial image-to-image style transfer is often used to bridge this gap, but it is not directly connected to the object detection task and can be unstable. We propose AWADA, a framework that combines attention-weighted adversarial domain adaptation connecting style transfer and object detection. By using object detector proposals to create attention maps for foreground objects, we focus the style transfer on these regions and stabilize the training process. Our results demonstrate that AWADA can reach state-of-the-art unsupervised domain adaptation performance in three commonly used benchmarks.