research overview

I develop algorithms for interpreting videos, especially video of outdoor cameras. In this domain, I solve a range of traditional computer vision problems, from calibration to tracking, using video of a scene captured over weeks, months, and even years.


Check the archives for old news.

selected projects

A false-color image displaying the correlation between a single pixel’s (in the blue box) time series and the time series of all other pixels in the image. This shows that nearby pixels are generally more correlated than distant pixels. using clouds to understand outdoor scenes - In this work, we answer the question, “What can shadows cast by passing clouds tell us about 3D scene geometry?”
The average face of people in different parts of the world. facial appearance around the world - attempting to quantify the relationship between human appearance and geographic location
webcam geolocalization - methods for estimating the location of a static outdoor camera using a collection of time-stamped images
webcam_montage global webcam archive - Tens of thousands of outdoor webcams are currently streaming live images from around the world. How can we use this imaging network to learn more about the world and how it changes over time?
streamlines using cloud motion to calibrate webcams - Calibrating a camera without being able to physically access or control the camera is a challenging problem. In this work, we show how video from a few partly cloudy days is sufficient to estimate the geo-orientation and focal length of an outdoor static camera.
Image of an urban scene with highlighted traffic lanes. tracking in structured scenes - Object tracking can be made much more efficient by exploiting the consistent patterns of motion (e.g. cars tend to travel along roads) found in the scene. Here we explore an extreme variant of this that simply discards all image data not along the center line of a lane of traffic.