Greg 2006 Material

Current Research

Current research:

• Caltech 256: Images and Downloadable Tar File
  • To do: make a prettier web page

• Pyramid Matching
  • Started as an EE148 project with Will Coulter (see our old project page)
  • This summer I'm trying to improve performance in various ways (see Summer 2006 Notebook)
  • This will hopefully be a good vehicle for testing some new projects (see below)

• The Fade Cascade is our attempt to allow decision cascade algorithms (e.g. Viola and Jones face detection) to learn with fewer training examples. This is a collaboration with Jeff Edlund.

• Cat detector (a project with Francois Fleuret)
  • Traditional human face- detection algorithms may fail in more complex situations
  • Example: cats pictures differ a great deal in pose, species, etc.
  • Scan real images: does Pyramid Matching achieve acceptable performance?

• Detection of Objects in Many (>100) Object Categories
  • I plan on exploting 3 different techniques to make this computationally feasible in realtime.
    • Coarse-to-fine object detection
      • Examples: Fleuret and Geman, Viola and Jones
      • Run coarse or simple detectors first to find areas that deserve more attention
      • Run fine or complex
    • Train on meta-categories, like "animal" or "tool" (new?)
      • Does it work?
    • Eigenmodels (new?)
      • Work out formalism
      • Tests

These topics are interrelated:

• Generalization of pyramid matching to large test images and many categories
• Categories we know vs. clutter vs. categories we don't know
• Training On A Hierarchy of Categories

Future Research

Topics that haven't been fleshed out yet:

• Diagonalizing the Confusion Matrix: EigenHypotheses
• Radially Binned Fourier Transforms: an alternative feature set to SIFT
• Exploiting semantic relationships to improve visual classification performance and benchmarking
• Why does bagging work? (an entropy-based explanation)
• Analysis of Klimt Paintings