Machine Learning and Friends Lunch

home
past talks
resources
conferences

Computer Vision and Machine Learning - Representations and Algorithms for Creating Interactive Visual Media

Abstract

Standard video is a simple stack of rectangular images of fixed resolution, displayed linearly in time. Can we create a representation of a scene that places dynamic, high resolution content in the directions we are most likely to look? I tell you how to do this.

I start with a top down view of how to decompose a complex scene using probabilistic generative models. I contrast the computational efficiency and representational power of different models. I then focus on the goal of efficiently constructing a high quality representation and re-examine the tasks involved. I outline the critical computations required for registering high-definition quality video frames into huge, high quality panoramic representations. I show how performing this task effectively and efficiently then facilitates the construction of rich, animated and interactive representations. Finally, I show how effective video registration also empowers an interactive system for real-time panoramic photography. This talk represents a summary of my contributions described in a series of papers at ECCV, ICCV, CVPR, SIGGRAPH, OZCHI and other venues.