Grassmannian graph-attentional landmark selection for domain adaptation

Domain adaptation aims to leverage information from the source domain improve classification performance in target domain. It mainly utilizes two schemes: sample reweighting and feature matching. While first scheme allocates different weights individual samples, second matches of domains using global structural statistics. The schemes are complementary with each other, which expected jointly work for robust adaptation. Several methods combine schemes, but underlying relationship samples is insufficiently analyzed due neglect hierarchy geometric properties between samples. To better advantages we propose a Grassmannian graph-attentional landmark selection (GGLS) framework GGLS presents attention-induced neighbors graphical structure performs distribution knowledge over Grassmann manifold. former treats landmarks differently, latter avoids distortion achieves properties. Experimental results on real-world cross-domain visual recognition tasks demonstrate that provides accuracies compared state-of-the-art methods.