About AMP Lab Projects Downloads Publications People Links
Porter Hall B10
Lab: Porter Hall B6
Department of ECE, Carnegie Mellon University,
5000 Forbes Avenue, Pittsburgh,
[Research Interests] [Project]
Video Coding & DSP Target Development
A robust/efficient face-eye tracking system is an essential component of a face recognition system. Based on colour distribution models and deformable template, we aim to make the system more robust to indoor lighting conditions.
A universal "face space" is typically of 200~400 dimensions is needed for eigenface-based face recognition. We propose a hierarchical approach to reduce computational requirement.
Top of this page
Face Recognition - Computational Load Reduction
Typical eigenface-based face recognition approach requires a face space of 200~400 dimensions. A query will involve computing 200~400 matrix calculations to derive the feature vector, and then identifying the best match in the feature database. This database can be viewed as a 1-level tree with N leaves (each leaf represents a class). The feature database can be represented as a binary tree; this transforms place a single N-class classification problem into m 2-class classification problems, with m ~ log (n). The computational requirement of matrix calculation is shown to be (m*d) < 200~400, where d is the dimension of face space needed to solve a 2-class problem.
Top of this pag