Calibrating cameras from internet images

Speaker

Aseem Agarwala

Date

Nov 1, 2007

Time

2:30PM to 3:30PM

Place

GRAIL (CSE 291)

Abstract

This talk will be a short and informal discussion on ongoing, half-baked research (with Dan Goldman, and Sujit Kuthirummal and Shree Nayar from Columbia) that we hope to get feedback on.

A specific camera model will have many particularities that cause it to deviate from the standard pinhole camera model used in computer vision; these include a non-linear response function, vignetting, radial distortion, chromatic aberration, and dead pixels. There are a multitude of techniques for calibrating cameras to measure these phenomena, but they are typically laborious and involve capturing many "special" images. We are exploring a different approach, that takes advantage of the fact that there are millions of images online in a database such as flickr.com that are labeled with their camera model and other details in their EXIF tags. Given such a dataset, can we learn the particularities of a camera by examining the statistics of thousands of images taken by it? And if so, can we calibrate the world's cameras by simply crawling the web, without requiring any access to the cameras themselves, any user effort, or any special calibration images? In this discussion I'll report on our initial efforts to measure camera vignetting, and mention a so-far un-tested approach to measuring camera response functions.

Graphics Seminar Home