The quality of a face recognition system heavily depends on the amount of labeled training data. Bias would probably exist in both in-class and between-class scatter when there is few labeled data. Considering  the cost of manual labeling is too high

an alternative choice is to make use of existing data which is related to the objective problem. In this way

it is able to alleviate the dependence of manual labeling via exploring numerous related data

offering a feasible solution to the case of lacking sufficient labeled training samples. To this end

a face transfer learning approach based on the sparse subspace clustering and robust principal component analysis is proposed

which allows employing unlabeled source data under the multi subspace assumption and mining useful information for the objective problem.