More about VLAD: A leap from Euclidean to Riemannian manifolds

Texture image classification with Riemannian fisher vectors

“…. , K by using the intrinsic k-means algorithm on a Riemannian manifold [7]. Thus, for each cluster c j , the cost function…”

“…In addition, a hard assignment scheme is applied. Starting from the definition of the elements v i in the R-VLAD descriptor [7]:…”

“…In other way, the VLAD descriptors can be obtained starting from FV, by taking into consideration only the derivatives with respect to the means of the GMM. Note also that the homoscedasticity assumption and the hard assignment scheme are required to obtain VLAD features [7,10].…”

“…In [3] and [17], the BoW approach has been extended to the so-called logEuclidean bag of words (LE-BoW) and bag of Riemannian words (BoRW) models by considering respectively the logEuclidean and geodesic distance between two points on the manifold. In addition, the Riemannian version of the VLAD method (R-VLAD) has been developed in [7] and has shown superior classification performances, compared to the classic VLAD algorithm.…”

“…These descriptors have multiple advantages. First, the obtained information can be used in a wide variety of applications like classification [2] and categorization [4], text [5] and image [6] retrieval, action and face recognition [7], etc. Second, combined with powerful local feature descriptors such as SIFT, they are robust to transformations like scaling, translation, or occlusion [7].…”