Charles T. Curry

Home Page ( at UCSC )

Picture of Charles T. Curry

Charlie Curry

University of California, Santa Cruz

Department of Applied Math and Statistics

PhD Student (entered 2003)

Advisor: Bruno Sanso

Email address: ctc@ams.ucsc.edu

Current Interests

  • Bayesian Statistics
  • Linear Models
  • Covariance Inference
  • Monte Carlo Computation
  • Computer Vision

    • Resume PDF

    Recent Papers

  • 2007 UCSC Master's Capstone: Inference for Climate System Properties PDF
  • 2005 UCSC Tech Report: Inference for Climate System Properties PDF

    • Recent Presentations

  • 2006 JSM Presentation: High-Dimensional Covariance Inference PPT
  • 2005 JSM Presentation: Inference for Climate System Properties PDF

    • The current state of my professional life.

    The current state of my academic life.

    I now know two different Cauchy jokes, one from Probability (thanks Thanasios) and one from Electrical Engineering (you too, Dileep). Why aren't there any good Gaussian jokes? Was he so mean?

    Current Collaborations

  • Covariance estimation model software tar, gzipped
  • Covariance estimation model figure Large PDF
    Problems with bias in the eigenvalues, a great deal of variability in the eigenvectors, and difficulty interpreting the posterior mean of each eigenvector.
  • Bayesian Case Studes Poster tar, gzipped
  • Comparison of eigenvalue posteriors under conjugate and separability assumptions PDF
  • 2005 CEF, BS, CTC: Scientific Paper Draft tar, gzipped

    • Home page(s) elsewhere: mac.com

    Vision Notes

    After many discussions of stereo video (Mark), laser range finding (Mike), information theory (Kevin K.), visual cortex modeling, textrons and Markov random fields (Kevin W.), I've decided to post my UCSC CMPS 260 (Computer Graphics Spring 2004) final project. Just a nice little example of how far a simple probability model, with minimal "ad-hockery" (as E.T. Jaynes might say), coupled with a sequential Monte Carlo sampling strategy can get you.

    MPEG4 Movie (QuickTime recording)

    The link is a screen capture (from Ambrosia Software's SnapzPro) of a face tracking application developed over one month in the spring of 2004. The app solves the inverse problem of locating and orienting a face model in 3D space given a single color video stream. Uses particle filters to track the posterior of the position and orientation parameters, and graphics hardware to evaluate the forward model and likelihood. In the video, you should be able to see the difference in the estimation error/noise under different numbers of particles (16 vs. 64 vs. 1024).

    All code was written from scratch in C++ (note the flexibility and robustness): leveraged OpenGL for texture polygon rendering, lighting and line drawing, the OpenGL shading pipeline for likelihood evaluation, Quicktime for video file loading and playback, and Apple's Cocoa user interface toolkit for the buttons, sliders, panels, trays, menus, etc.

    Couple these techniques with a good covariance structure and the additional info provided by a 3D vision system (like stereo or laser sensors) and you could almost certainly do even better.

    I apologize if watching the movie bores you to tears: it wasn't really rehearsed. I just exercise the features as I remembered them (it's been over a year since I last used the app). One really neat part to watch is the sequence of steps from "Sample Prior" to "Update Likelihoods" to "Importance Resample". Note the way all the particle configurations snap to a small number of really good estimates.