VDM-DA: Virtual Domain Modeling for Source Data-Free Domain Adaptation

Domain adaptation aims to leverage a label-rich domain (the source domain) help model learning in label-scarce target domain). Most methods require the co-existence of and samples reduce distribution mismatch. However, access may not always be feasible real-world applications due different problems ( e.g. , storage, transmission, privacy issues). In this work, we deal with data-free unsupervised problem propose novel approach referred as Virtual Modeling for Adaptation (VDM-DA), which virtual acts bridge between domains. Specifically, based on pre-trained model, generate by using an approximated Gaussian Mixture Model (GMM) feature space, such that maintains similar without original data. Moreover, also design effective alignment method divergence gradually improving compactness through learning. way, successfully achieve goal domains when training deep networks We conduct extensive experiments four benchmark datasets both 2D image-based 3D point cloud-based cross-domain object recognition tasks, where proposed (VDM-DA) achieves promising performance all datasets.