MULTICLASS DOMAIN ADAPTATIONWITH ITERATIVE MANIFOLD ALIGNMENT

We propose a novel approach for multiclass domain adaptation using an iterative manifold alignment technique inspired by the TRiplet-based Iterative ALignment (TRIAL) protein structure alignment algorithm. Our technique learns a rigid transformation for each class using a set of automatically-selected pivot samples that char-acterize the relative relationships between classes in two similar, but not identical, feature spaces. We demonstrate that our technique robustly reconciles domain-specific differences between similar classes in hyperspectral images captured under different conditions, and yields more accurate results than recently-proposed manifold alignment techniques. We evaluate our method on a pair of real-world hyperspectral images of Cuprite, NV, and provide a MATLAB implementation of our algorithm, available online. Index Terms — domain adaptation, manifold alignment, multi