In this paper we propose an information-geometric method for comparing superpixel (turbopixel) images. Turbopixels are encoded by tensors and they are referred to as TurboTensors. Our methodology has three ingredients. Firstly, we formulate the comparison of the turbopixels topology in terms of the non-rigid alignment of the Isomap embedding of the weighted adjacency matrices; we propose a multi-dimensional information-theoretic dissimilarity measure. Secondly, we formulate the comparison of bags-of-turbopixels through tangent spaces de-projection and multi-dimensional and non-parametric information-theoretic dissimilarity measures. Thirdly, we combine the two latter elements into a flexible energy function whose minimization yields the optimal matching of superpixels images as well as their similarity. In our experiments we show that the proposed method is a useful tool for finding clusters in image sequences. Finally, we show that our approach outperforms state-of-the-art image comparison through non-rigid and affine matching of SURF features. © 2013 Springer-Verlag.
CITATION STYLE
Escolano, F., Hancock, E. R., Bonev, B., & Lozano, M. A. (2013). TurboTensors for entropic image comparison. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 7877 LNCS, pp. 51–60). https://doi.org/10.1007/978-3-642-38221-5_6
Mendeley helps you to discover research relevant for your work.