Pose and appearance based clustering of face images on manifolds and face recognition applications thereof
This dissertation takes a small but important step towards solving the following general problem of current interest: Assuming that each individual in a population can be modeled by a single frontal RGBD face image, is it possible to carry out face recognition in the wild for such a population? Face recognition in the wild means basing the identity decision on a set of images captured by a network of cameras, with each camera viewing the face from an arbitrary viewpoint. The general problem as stated above is extremely challenging. It, however, throws up subproblems that can be addressed today. The subproblems addressed in this dissertation relate to: (1) Generating for each individual a large set of viewpoint dependent face images from a single RGBD frontal image; (2) Discovering through a comparative evaluation the best algorithm among ISOMAP, LLE, and LPCA to use for clustering the viewpoint-dependent model data so generated on the manifolds on which it resides; (3) Comparing a global approach to classification in which all of the training data resides in the same subspace with hierarchical approaches based on view-partitioned subspaces for representing the training data; and (4) Using a weighted voting algorithm for integrating the evidence collected from multiple images of the same face as recorded from different viewpoints. For the results shown, the frontal RGBD image of each individual in a collection of 10 people is used to generate 925 viewpoint dependent face images. Although this constitutes a small population, our results nonetheless provide important insights for further extensions of this research.
Kak, Purdue University.
Computer Engineering|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our