Unified Framework for Visual Domain Adaptation Using Globality-Locality Preserving Projections

摘要

Domain Adaptation is a segment of machine learning that allows us to learn from a labelled source data distribution to classify different but related unlabelled target data distribution. In this paper, we propose a novel framework called Unified Framework for Visual Domain Adaptation using Globality-Locality Preserving Projections (UFVDA) that reduces the divergence between source and target domain both statistically and geometrically. In this framework, we use Globality-Locality Preserving Projections (GLPP) instead of primitive methods such as Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA) for dimensionality reduction and two projection vectors to project the source and the target domain data onto a common subspace. The better performance of our proposed framework than other state-of-the-art visual domain adaptation and the primitive dimensional reduction methods on real-world domain adaptation data-sets has been verified by extensive experiments. Our proposed approach UFVDA achieved a mean accuracy of 84.09% and 79.35% for all tasks of Office-Caltech data-set with VGG-Net features and PIE Face Recognition data-set respectively.