MIC: Masked Image Consistency for Context-Enhanced Domain Adaptation

In unsupervised domain adaptation (UDA), a model trained on source data (e.g.synthetic) is adapted to target data (e.g. real-world) without access to targetannotation. Most previous UDA methods struggle with classes that have a similarvisual appearance on the target domain as no ground truth is available to learnthe slight appearance differences. To address this problem, we propose a MaskedImage Consistency (MIC) module to enhance UDA by learning spatial contextrelations of the target domain as additional clues for robust visualrecognition. MIC enforces the consistency between predictions of masked targetimages, where random patches are withheld, and pseudo-labels that are generatedbased on the complete image by an exponential moving average teacher. Tominimize the consistency loss, the network has to learn to infer thepredictions of the masked regions from their context. Due to its simple anduniversal concept, MIC can be integrated into various UDA methods acrossdifferent visual recognition tasks such as image classification, semanticsegmentation, and object detection. MIC significantly improves thestate-of-the-art performance across the different recognition tasks forsynthetic-to-real, day-to-nighttime, and clear-to-adverse-weather UDA. Forinstance, MIC achieves an unprecedented UDA performance of 75.9 mIoU and 92.8%on GTA-to-Cityscapes and VisDA-2017, respectively, which corresponds to animprovement of +2.1 and +3.0 percent points over the previous state of the art.The implementation is available at https://github.com/lhoyer/MIC.