Probabilistic Graphical Models For Visual Feature Analysis

There is a trend of utilizing local visual features to solve object recognition prob-lem in computer vision. The key point is how to make good use of local visual features,and extract semantics in the images accurately. On the other hand, probabilistic graph-ical models from the area of machine learning make several successes in various fields.This paper describes the theory and algorithms of probabilistic graphical models, andits applications in computer vision. Moreover, this paper extends LDA, a model fromtext mining area, rendering it more suitable for vision problem.This paper firstly makes a concise survey on modeling in computer vision, withan emphasis on models belonging to the framework probabilistic graphical models.Besides, this paper reviews the the theory and algorithms of probabilistic graphicalmodels, especially Bayesian networks, EM algorithm, variational inference and Gibbssampling. Then LDA and its inference and learning algorithms are elaborated andanalyzed. Finally, this paper proposes an extension of LDA, affine invariant topicmodel(AITM), and applies AITM to object recognition. Experiments on two datasetsdemonstrate AITM's efficiency.The main contribution of this paper is AITM. As an extension of LDA, AITMabandons the"bag of words"assumption in LDA and incorporates spatial structureof visual features. Not modeling the locations of local features directly, AITM assumesthat the location of local features is affinely transformed from the prior location, andmodels prior locations and affine transforms as latent variables. The merit of AITMis that AITM is able to describe spatial structure of visual features in relatively fewparameters, avoiding the exponentially rapid increase due to feature combination.