Large-scale semi-supervised learning by Approximate Laplacian Eigenmaps, VLAD and pyramids

The paper builds upon recent advances in feature representation and dimensionality reduction to propose a semi-super-vised image annotation framework that achieves state-of-the-art accuracy at substantial gains in computation cost. More specifically, the framework combines the VLAD feature aggregation method with spatial pyramids and PCA for image representation, and proposes the use of Approximate Laplacian Eigenmaps (ALEs) for learning concepts in time linear to the number of images (labeled and unlabeled) available at training. A set of thorough experiments on MIR-Flickr and ImageCLEF 2012 ground truth annotations explore the impact of PCA and pyramids on the attained accuracy, and demonstrate that the proposed framework achieves virtually the same accuracy with a state-of-the-art manifold learning approach, while at the same time offering substantial speedup (in the order of ×80) making possible the completion of a training/testing run for a set of 25k images in less than 3 minutes in a commodity workstation.