In a class project I reimplemented the Grabcut algorithm. I discovered a few corrections to the original equations and released our source code.
GrabCut is an innovative 2D image segmentation technique developed by Rother et al. . This paper provides implementation details omitted from the original paper. Details covered in background papers are summarized here so that future implementors can refer to a single paper. Our implementation of GrabCut is described and results are included. Our main contribution is correcting errors in Equations (9) and (11) of the original paper. We also discuss weaknesses of the algorithm that were not discussed in the original paper. We present possible research directions to address these problems.
Workstation Idle Period Duration Prediction for Distributed Computing
2004. This was a class project that used Bayesian methods to predict idle periods in a distributed computing platform.
We consider the problem of predicting idle period duration on workstations used in DOGMA, a distributed computing architecture developed and deployed at Brigham Young University. BYU provides students with open access to nearly 900 workstations throughout the main campus for use in homework and research pursuits. When idle, these computers are used by DOGMA to perform time consuming research computing. Predicting the lengths of future idle periods could improve the performance of DOGMA. We develop a hierarchical Bayesian model for idle period lengths. We then use Markov Chain Monte Carlo (MCMC) computation to create an idle duration predictive distribution for each computer. We demonstrate that a Bayesian approach produces reasonable predictive distributions
for this application.