Stepwise Domain Adaptation (SDA) for Object Detection in Autonomous Vehicles Using an Adaptive CenterNet
Guofa Li, Zefeng Ji, Xingda Qu
Abstract
In recent years, deep learning technologies for object detection have made great progress and have powered the emergence of state-of-the-art models to address object detection problems. Since the domain shift can make detectors unstable or even crash, the detection of cross-domain becomes very important for the design of object detectors. However, traditional deep learning technologies for object detection always rely on a large amount of reliable ground-truth labelling that is laborious, costly, and time-consuming. Although an advanced approach CycleGAN has been proposed for cross-domain object detection tasks, the ability of CycleGAN to reduce the divergence across domains at the feature level is limited. In this paper, a stepwise domain adaptation (SDA) detection method is proposed to further improve the performance of CycleGAN by minimizing the divergence in cross-domain object detection tasks. Specifically, the domain shift is addressed in two steps. In the first step, to bridge the domain gap, an unpaired image-to-image translator is trained to construct a fake target domain by translating the source images to the similar ones in the target domain. In the second step, to further minimize divergence across domains, an adaptive CenterNet is designed to align distributions at the feature level in an adversarial learning manner. Our proposed method is evaluated in domain shift scenarios based on the driving datasets including Cityscapes, Foggy Cityscapes, SIM10k, and BDD100K. The results show that our method is superior to the state-of-the-art methods and is effective for object detection in domain shift scenarios.