Modeling Sequential Domain Shift through Estimation of Optimal Sub-spaces for Categorization
In Proceedings British Machine Vision Conference 2014
AbstractThis paper describes a new method of unsupervised domain adaptation (DA) using the properties of the sub-spaces spanning the source and target domains, when projected along a path in the Grassmann manifold. Our proposed method uses both the geometrical and the statistical properties of the subspaces spanning the two domains to estimate a sequence of optimal intermediary subspaces. This creates a path of shortest length between the sub-spaces of source and target domains, where the distributions of the projected source and target domain data are identical when projected onto these intermediate sub-spaces (lying along the path). We extend our concept to the kernel space and perform non-linear projections on the subspaces using kernel trick. Projections of the source and target domains onto these intermediary sub-spaces are used to obtain the incremental (or gradual) change in the geometrical as well as the statistical properties of sub-spaces spanning the source and target domains. Results on object and event categorization using real-world datasets, show that our proposed optimal path in the Grassmann manifold produces better results for the problem of DA than the usual geodesic path.
FilesExtended Abstract (PDF, 1 page, 109K)
Paper (PDF, 12 pages, 315K)