Thesis Presentation

We introduce Peekaboom, a web-based game built to gather large datasets for computer vision applications. Peekaboom is an instance of a human computation algorithm -- in other words, it incorporates the behavior of humans in such a way that, just by playing the game, people teach computers the locations of objects in images. We begin by describing the game and its features, which passively ensure the correctness and quality of collected data. Then, we show how Peekaboom produces its intended output of object location information through qualitative and quantitative evaluation in the form a user study and a computer vision experiment. Lastly, we discuss future application areas for the Peekaboom dataset. This talk is meant for a general purpose audience, so no technical background is necessary.

Thesis Committee:
Manuel Blum, Co-Chair
Luis von Ahn, Co-Chair