Online Domain Adaptation for Semantic Segmentation in Ever-Changing Conditions

AbstractUnsupervised Domain Adaptation (UDA) aims at reducing the domain gap between training and testing data is, in most cases, carried out offline manner. However, changes may occur continuously unpredictably during deployment (e.g. sudden weather changes). In such conditions, deep neural networks witness dramatic drops accuracy adaptation not be enough to contrast it. this paper, we tackle Online (OnDA) for semantic segmentation. We design a pipeline that is robust continuous shifts, either gradual or sudden, evaluate it case of rainy foggy scenarios. Our experiments show our framework can effectively adapt new domains deployment, while being affected by catastrophic forgetting previous domains.