Representing dynamic textures based on polarized gradient features
Résumé
Efficiently representing dynamic textures (DTs) is one of the significant challenges for video understanding in real implementations of computer vision applications. In this work, an efficient approach for DT description is introduced by addressing the following prominent concepts. Firstly, high-order 2D/3D Gaussian-gradient filtering kernels are used for filtering a given video to obtain its Gaussian-gradient-filtered images/volumes. Secondly, taking advantage of the polarizing property of these responses, we propose a competent model of decomposition to decompose them into corresponding robust collections of separately polarized outcomes, which are complementary to DT representation. Finally, a simple variant of local binary patterns (LBPs) is applied to extract local polarized Gaussian-gradient features from the complemented collections for constructing discriminative local-based descriptors. Experimental results for DT recognition on benchmark datasets have remarkably validated the efficacy of our proposal.
Origine : Fichiers produits par l'(les) auteur(s)