The efficiency of covariance descriptors (CovDs) has been explored in several image/video categorization tasks. CovDs lie on Riemannian manifolds known as tensor manifolds. Therefore, the non-Euclidean geometry should be taken into account in devising inference methods that exploit them. In this chapter, we extend the conventional bag-of-words model from Euclidean space to non-Euclidean Riemannian manifolds. To this end, we elaborate on an intrinsic bag-of-Riemannian-words (BoRW) model, which takes into account the true geometry of tensors in obtaining its codebook and histogram. Experiments on challenging a virus texture data set show that the proposed BoRW on CovDs obtains notable improvements in discrimination accuracy, in comparison to popular bag-of-words models.
|Title of host publication||Case Studies in Intelligent Computing|
|Subtitle of host publication||Achievements and Trends|
|Editors||Biju Issac, Nauman Israr|
|Place of Publication||New York USA|
|Number of pages||13|
|Publication status||Published - 2014|