3D Multi-Views Object Classification Based on a Fully Generalized Dirichlet Allocation Model

Ahmed Yasser Eita, Hafsa Ennajari, Nizar Bouguila

Abstract: Comparing 3D objects based on their local features necessitates a substantial amount of computational resources. Recent studies have demonstrated that the combination of topic modeling and the Bag of Visual Words (BoVWs) approach can effectively capture useful and distinguishable object rep-resentations in a consistent way. One of the most common topic modeling approaches is LDA which is based on Dirichlet distribution priors, but it is limited by its incapacity in modeling topic correlations. Consequently, several extensions of LDA were proposed to solve this problem including GD-LDA, LGDA, and CVB-LGDA which have shown good results in discovering the semantic relationships between topics but are either suffering from incomplete generative processes assumptions that impact their inference efficiency or require high-computational power due to their complexity. In this paper, we introduce F-GDA, a fully Generalized Dirichlet Allocation model that is mainly derived from Generalized Dirichlet distributions for 3D objects recognition. Unlike GD-LDA which generalizes only the topics parameter, F-GDA generalizes all the model priors parameters, ensuring complete flexibility in the priors. Extensive experimental results have demonstrated the ability of our model to learn high-quality data representations of a real-world 3D Multi-views dataset ETH80 and also N15.