Research Interests
I am interested in how end-to-end vision systems -- both machine and biological -- are put together. Although the fields of computer vision and cognitive science continue to make progress in understanding parts of the visual process, we still have only sketchy notions of how these parts are integrated. At the moment, I am focusing on building a computational analogue to the human ventral visual system, since it is the ventral system (and not the dorsal or early visual systems) that seems to correspond most closely with our intuitive notions of visual perception. My approach is to build an end-to-end computer vision system that approximately implements the psychological models proposed by Stephen Kosslyn (with due note of Milner & Goodale, others), and to use this model to provide feedback to the larger cognitive science community about its adequacy. This project is called Modeling the Ventral Visual Pathway: A Biomimetic Approach to Object Recognition. Ross Beveridge is a co-PI.
In addition to modeling the human ventral visual pathway, I work as a co-PI on a HumanID project doing statistical evaluations of face recognition algorithms. The goal here is to directly compare face recognition algorithms such as principal component analysis (PCA), linear discriminant analysis (LDA), independent component analysis (ICA), and Gabor jet matching. The analysis is not just a "bake off", however; we want to understand the strengths and weaknesses of each approach. Ross Beveridge is the PI for this project (I'm a co-PI), which is called Evaluation of Face Recognition Algorithms and is funded by DARPA. Since some form of subspace projection algorithm would seem to be a likely component of Kosslyn's hypothesized exemplar categorization subsystem, this work is complimentary with my interest in biological modeling.
My other funded research is as part of the Cameron project. The Cameron project attempts to compile sequences of image processing procedures written in a high-level computer language directly to circuits in an adaptive computing system (a.k.a. reconfigurable computing system). Our goal in the Cameron project is to develop a high-level language and optimizing compiler so that image processing applications on adaptive computing systems can be programmed at a software (rather than circuit design) level. Wim Bohm is the PI on this DARPA funded project, with Ross Beveridge and myself as co-PIs.
Although currently unfunded, the Adaptive Object Recognition (ADORE) project seeks to learn task-specific object recognition strategies from examples. The underlying intuition is that object recognition can be modeled as a Markov Decision Problem (MDP), and that control policies can be trained to sequence vision procedures in order to recognize specific objects in known domains.
In the past I have also worked on color analysis in outdoor images, persistent data stores for computer vision, and visually-guided autonomous outdoor vehicles.