Robust Linear Subspace for Image Set Retrieval

This paper attempts to take advantage of both dual linear regression and sparse coding for set-to-set based object recognition. In order to determine the right category of test image set, our algorithm finds a virtual object in the intersection of two subspace, one of which is represented by a sparse linear combination of the images in test image set, the other represented by a sparse linear combinations of all images in the gallery. The quality of the representation of the virtual object using images of each category in the gallery is evaluated and used to make the decision of the classification. Experiments on the benchmarks Caltech101, YouTube and LFW are carried out to verify the effectiveness of the algorithm. The results demonstrate that our algorithm achieved best classification accuracy with state-of-the-art methods.