Saliency Detection via Sparse Reconstruction Errors of Covariance Descriptors on Riemannian Manifolds

Abstract

We present a novel visual saliency detection method using covariance matrices on a Riemannian manifold. After over-segmentation, superpixels are generated and featured by the region covariance matrix. The superpixels on image boundary are regarded as possible background cues and are used to build the background dictionary. A sparse model is then constructed based on the background dictionary, where a kernel method, embedding Riemannian manifolds into reproducing kernel Hilbert space, is used. For each superpixel, we compute sparse reconstruction errors as a saliency measurement, which are then weighted based on the local context and global context information. Finally, multi-scale reconstruction errors are integrated to reduce the effect of the scale problem, and an object-biased Gaussian model is adopted to refine the saliency map. The main contribution of this paper is using a kernel sparse representation of the region covariance descriptors for saliency detection. Experiments with public benchmark dataset show that the proposed algorithm outperforms the state-of-the-art methods in terms of precision, recall, and mean absolute error, which demonstrate that our method is more effective in uniformly highlighting salient objects and is robust to background noise.