Disentangled Discriminator for Unsupervised Domain Adaptation on Object Detection

摘要

Object detection plays an important role in computer vision tasks such as autonomous driving, robotics, etc. Typically, a detection model is firstly trained on collected data and then deployed in real world. However, the discrepancy exists between training (source) and testing (target) data, which degrades the detection model's performance in the real world. To mitigate the negative effects, Unsupervised Domain Adaptation (UDA) methods learn the features of a shared domain via a discriminator. However, existing discriminators consider only the in-distribution adversarial learning, which ignore the out-of-distribution data of individual domains. In this paper, we propose a disentangled discriminator to consider the in-distribution and outliers separately. It aligns the source and target data with split branches under a gated strategy. We combine the disentangled discriminator with a Teacher-Student (T-S) framework that trains the student using labeled source data and unlabeled target data under a self-training mechanism. Specifically, the teacher network, that is updated with the parameters of student network via the exponential moving average, predicts pseudo labels for unlabeled data. The quality of pseudo labels can be improved after alleviating the domain discrepancy thanks to the disentangled discriminator. Extensive experiments on benchmarks demonstrate the superiority of the proposed method. Specifically, we achieve 53.9% mAP on Foggy Cityscapes, which is 7.2% higher than the Oracle.