Paul Ardis
main
life
research
seminar
department
publications
contact
index » research
Probabilistic Modeling of Human Visual Attention
a test image  a corresponding visual interest map

Human observer interest can be tractably described using graphical model-motivated energy minimization problems, with large existing eyetracking datasets used for training. The effects of well-studied components such as Persistence of Vision, Foveal Coverage, Inhibition of Return, Local Contrast, and Covert Attention can be estimated using fast probabilistic search methods in order to simulate V1 neuronal processing, with comparison provided against leading psychophysical models (based upon more complex predictive structures).
[ Temporarily Removed due to Double-Blind Publication Constraints ]
Feature-Based Video Registration
an image of trajectories used for video alignment

I have explored systematic advances in automatic spatio-temporal registration of consumer video, including:
  • Sub-sequence voting schemes
  • Relaxation of feature uniqueness
  • Descriptor "appearance" priors
  • Trajectory sampling
This work was largely motivated by various publications (CVPR2000, ICCV2001, PAMI2002, CVPR2004, PAMI2005) of Michal Irani at the Weiszmann Institute in Rehovot (Israel), and was part of an ongoing collaboration between Kodak Research Laboratories and the University of Rochester.
Salience-Guided Graphical Modeling for Inpaint Method Selection
Taking as input a correlation of salience (visual cortex excitation) with image or video data, it is possible to learn an optimal decision boundary relating salience scores and the use of fast interpolation methods over slower patch-based methods that guarantee good structure continuation. This boundary is determined by way of feature function weighting in a template-based Conditional Random Field, where decision boundary construction approximates inference estimation with calculable error bounds and decision regret. More general connections between probabilistic graphical model theory and decision theory are also pursued on a theoretic level.
... and more