Estimating the speed of a passing object; estimating the distance between two objects in a scene far two objects are far each other; refocusing your eye from near to far… You perform these tasks more than 100,000 times per day.  The effortlessness with which do these things belies the enormous computational complexity required to do them them well. Our work concentrates on understanding how to estimate depth from natural images. We determine how best to estimate individual depth cues from natural images (e.g. defocus, disparity, motion) using mathematical and computational methods. We determine how well humans estimate those same cues using behavioral (psychophysical) experiments.

A fundamental goal of vision research is to understand how vision functions in natural conditions with natural stimuli. How do we see? What are the computations that optimally transform sensory information into behaviorally relevant representations of the environment? What are the computations that humans and animals actually use?  Unfortunately, natural stimuli are monstrously complicated and are difficult to characterize mathematically. As a result, most vision research uses simple, artificial stimuli that are easier to characterize (e.g. bars and blobs). Most of our knowledge about visual processing derives from research with such stimuli; however, these stimuli lack many of the properties inherent to natural stimuli, the stimuli the visual system evolved to process.

The primary aim of our research is to enable the principled study of critical visual tasks with natural stimuli. Rather than attempting to develop a general model natural stimuli, we narrow the problem by focusing on the properties of natural stimuli that are most useful for particular tasks. We develop tools to enable rigorous mathematical characterization of task-relevant properties of natural stimuli. These tools help generate principled, quantitative hypotheses about how visual information should be ideally processed. These hypotheses are then used to design experiments testing behavioral performance and neural processing. In some cases, we have discovered a striking correspondence between ideal and human visual systems. Methods that are developed for the study of a given task in the human visual system can oftentimes be applied to a similar task in animal or machine vision systems.

Name the animals! Answers are upside down:

  1. Talks

  1. Research

  1. Teaching

  1. Music

  1. Patents

  1. People

  1. Code

  1. Projects

  1. Demos

  1. Press

  1. Contact

  1. CV