This page provides a series of suggested projects. Students are certainly free to suggest alternative project topics. However, I am offering these suggestions for several reasons.
First, In the past I've seen students have understandable difficulty designing projects early in the semester, before they have actually seen most of the material that will be covered. In other words, by the end of the semester, most of you will know enough to make informed choices as to what sort of project you might design, but unfortunately, by then it is much too late. A second reason for making these suggestions is that there are already resources and some background in place here for these projects. Many relate directly to prior and ongoing research conducted here and consequently are more likely to lead to significant results in the course of one semester. For example, we have on site on of the worlds largest and best studied sets of human faces: the FERET data set.
Most of the current research on human face recognition assumes frontal images which have been geometrically normalized. Typically this means the centers of the eyes have been found automatically or by hand, and then the image resampled so the eyes fall at known pixel coordinates. This is a weak link in the current approach to face recognition, since failure to performance normalization correctly will cause subsequent recognition failures. Moreover, automated normalization algorithms are less well developed than the corresponding matching algorithms intended for use after normalization.
This project could take several directions, but a general outline might go as follows. Phase 1 of the project could involve a literature survey of existing methods used to solve this problem, along with possible implementation of one method (presuming that method is relatively straight forward). Phase 2 would address some question of performance relative to one or more method. For example, the question under consideration might be: "Algorithm A performs normalization to within a tolerance of epsilon for 95% of images." Alternatively, recognition algorithm B performs face recognition as well using automated normalization algorithm A as when using hand normalized imagery."
One of the recurring question in the face recognition community is what, if anything, is so special about the PCA subspace for human face identification. The PCA subspace is only one of an infinite set of possible orthogonal basis sets. In work carried out here last semester, Professor Kirby in the Mathematics Department and I became curious what might happen if a series of randomly selected orthogonal basis sets were compared to the PCA set. Phase I of this project would involve first understanding the problem; providing a succinct statement of the problem and development of an algorithm for choosing random or psuedo-random alternative basis sets. Phase II would then address a questions such as: "Out of a large set of randomly chosen basis sets, with high likelihood one will arise that is superior to the PCA set for discriminating a given set of human faces."
Perhaps the best set of algorithms for optimally matching 2D object models, expressed as a set straight line segments, to noisy and cluttered image data has been developed here at CSU. This line of research has built upon my own Ph.D. work at the University of Massachusetts. We now have a mature system, LiME, which is freely distributed on the world wide web and is beginning to be used by others. In particular, it was recently used by a company involved in inspection of integrated circuits to register images.
There are a series of open topics concerning LiME. One is to better characterize the relative performance of the three primary algorithms currently implemented. Another is to refine one of these algorithms using more recent ideas associated with something called niching in the genetic algorithms community. Phase I in either case would involve becoming familiar with the existing LiME system, which is itself both a rich an complex environment. Then, were the goal to better understand the existing algorithms, phase II might address a question of the form: "For problem domain X, the Messy GA more reliably finds the correct match than does either local search or key-feature matching." There are already a series of publications relating to this question. However, I am still not entirely satisfied with these studies and additional work is required before I am ready to prepare a major journal article for the computer vision community on this topic. The ground work for such an article might be one practical out growth of this project.
Were the goal instead to develop a more refined, and potentially more expedient algorithm, then phase II would need to address a question of the form: "For domain X, new algorithm A finds better matches then any of the previous algorithms." Were this to proof true, the result would be exciting, since it is my believe that the current algorithms are already the best known for reliably solving these problems. While there is not space to go into detail, as suggested above, there are already a number of ideas in place for such an algorithm relating to niching strategies in genetic algorithm research, and this project could potentially involve considerable collaboration with Darnel Whitely and his student Jean-Paul Watson.
Baback Moghaddam and Alex Pentland at MIT developed an algorithm that measures similarity between images in terms of the a posterior probability of the differences between the two images being characteristic of images of the same person versus different people. This approach requires estimating the probability of the observed image difference, conditioned upon the image pair being of the same person. It also requires estimating the probability of the observed image difference conditioned upon the pair being images of different people. Phase I of this project would involve becoming familiar with this technique and developing an initial trial implementation. Phase II would then further test the implementation, seeing that it performed comparably with the original: this is possible since we have the same data. It would then address a question such as: "The MIT algorithm is superior to standard variants of PCA." In essence, this would designing and conducting experiments in an attempt to duplicate results observed in the original NIST FERET evaluation. Should this project go well, it would be even more interesting to consider questions relating to the broader question of why this algorithm works well.
W. Zhao, A. Krishnaswamy, R. Chellappa, D.L. Swets and J. Weng at the University of Maryland have developed an algorithm that uses a combination of Principal Components Analysis (PCA) and Linear Discriminant Analysis (LDA). Traditional PCA algorithms perform nearest neighbor matching in a subspace defined by the eigenvectors of a covariance matrix produced from a set of images, typically but not always a set containing examples of the people to be recognized. By using the k eigenvectors associated with the k largest eigenvalues, the subpace is tailored to preserve as much information about the raw variation between images as possible using only k orthogonal basis vectors. We have a similar algorithm already implemented by Wendy Yambor. Phase I of this project would more carefully review the exact algorithm proposed by Zhao et. al. and modify our implementation to match theirs. Phase II would then address the question: "The Maryland algorithm is superior to standard PCA under conditions X." Exploring and coming to understand what these conditions are would be one of the more significant aspects of this project. Indeed, a thorough understanding might well exceed the what can be expected in a semester. However, a good start could certainly be made.
As already described in the suggested project on Local Search and Messy GA Matching, we have a mature system for performing optimal 2D line segment matching. An intriguing question, and a somewhat unusual one, is whether this tool might be used to normalize face images to a canonical view. There are two parts to this question. One is wether one or several "average" models of the edge structure of human faces might be constructed. Second, could matches between these standard "average" faces and novel images provide reliable registration information for normalizing the novel image to a standard coordinate system. Phase I of this project would involve becoming familiar with LiME and at least one straight line segment extraction algorithm. Phase II would address the questions just posed: can we develop canonical models of faces, and are matches to these then a good basis for registration. This later question can be posed more formally as: "Performance of face recognition algorithm A using LiME normalized imagery is equivalent to performance using hand registered images."
[TOP]
Last Updated, 8/20/2000, copyright, J. Ross Beveridge.