3D Generic Elastic Models
Prabhu, U., Heo, J., and Savvides, M.
Unconstrained Pose-Invariant Face Recognition Using 3D Generic Elastic Models
IEEE Transactions on Pattern Analysis and Machine Intelligence vol.33, no.10, pp.1952-1961, Oct. 2011
author={Prabhu, U. and Jingu Heo and Savvides, M.},
journal={Pattern Analysis and Machine Intelligence, IEEE Transactions on},
title={Unconstrained Pose-Invariant Face Recognition Using 3D Generic Elastic Models},

Generating a 3D model of an object from a single 2D image is an inherently underconstrained problem. In order to be able to get a single solution, we must make some assumptions about the object that we are looking at. In the case of faces, we have seen from studying many 3D scans that the z coordinate of a point on the face seems to be fairly stable across people, at least wen broken down into ethnic groups. In other words, the left corner of the left eye tend to be very close in depth relative to the tip of the nose across people. By localizing a set of landmarks on the face, we can warp the depth image to the 2D image we see and generate the texture for the 3D model. However, a sparse set of landmarks is not enough to generate the 3D model's texture. We apply a Delaunay triangulation to the points and use loop sub-division to generate a denser set of correspondences. From this denser set, we can extract enough texture information to generate the 3D model of the face. The 3D models allow us to do many things with the faces such as synthesizing new poses of the face and performing recognition across poses very accurately.


Spartans: Single-sample Periocular-based Alignment-robust Recognition Technique Applied to Non-frontal Scenarios
F. Juefei-Xu, K. Luu, and M. Savvides
IEEE Transactions on Image Processing (TIP), vol. 24, no. 12, pp. 4780-4795, Dec. 2015.
author={Juefei-Xu, F. and Luu, K. and Savvides, M.},
journal={IEEE Trans. on Image Processing},
title={{Spartans: Single-sample Periocular-based Alignment-
robust Recognition Technique Applied to Non-frontal


These 3D models have been shown to be very accurate, even when using small regions of the face such as the periocular region. These small regions can be synthesized at different views to train new classifiers that have been shown to be outperforming other state-of-the-art methods.