Video-Based Face Recognition Using Bayesian Inference Model

Video-based Face Recognition using Spatio-Temporal Representations

“…Cluster centers are then selected as exemplars and a probabilistic voting strategy is used to classify new video sequences. Later exemplar-based works such as (Fan et al, 2005;Liu et al, 2006) performed classification using various Bayesian learning models to exploit the temporal continuity within video sequences. Liu et al (Liu et al, 2006) also introduced a spatio-temporal embedding that learns temporally clustered keyframes (or exemplars) which are then spatially embedded using nonparametric discriminant embedding.…”

“…In many previous works (Fan et al, 2005;Hadid & Peitikäinen, 2004), k-means clustering is the primary choice for assigning data into different clusters due to its straightforward implementation. However, it has some obvious limitations -firstly, it is sensitive to the initial seeds used, which can differ in every run, and secondly, it produces suboptimal results due to its inability to find global minima.…”

“…In this work, we propose a new classification method for video sequences using an exemplar-driven Bayesian network (EDBN) classifier, which introduces a joint probability function that is estimated by maximum a posteriori (MAP) decision rule within a Bayesian inference model. In comparison to other Bayesian methods (Fan et al, 2005;Liu et al, 2006), the EDBN classifier incorporates temporal continuity between successive video frames with consideration for the causal relationships between exemplars and their parent classes.…”

Learning Neighborhood Discriminative Manifolds for Video-Based Face Recognition

“…This limits their potential usage for classification and recognition tasks. Several works [7,8] reported good recognition rates in a video-based face recognition setting by using LLE to build a view-based low-dimensional embedding for exemplar selection, but stops short of utilizing it for feature representation.…”

“…For each cluster, the face that is nearest to the cluster mean is selected as an exemplar, or a representative image of the cluster. Similar to these approaches [7,8], the subject in each training video is represented by a set of M exemplars, which are automatically extracted from the video. Features will be extracted from the exemplar sets using the NDMP method.…”