| Cameras are a useful source of input for many interactive applications, but computer vision programming is difficult and requires specialized knowledge that is out of reach for many application developers. This dissertation contributes a novel architectural framework for developing computer vision applications, in which developers use visual examples to train and tune computer vision recognizers. We claim that this framework allows people with minimal programming experience, and no specialized knowledge of computer vision, to use camera data in a wide variety of compelling camera-based applications: home monitoring and surveillance, automated reminder systems, robotic navigation, interactive art, gesture-controlled appliances, video games, and many others.;We demonstrate our claim through the design and evaluation of a prototyping tool for camera-based interaction, which we built, distributed, and incrementally refined over the course of several years. In addition to presenting a new software framework for building computer vision applications, and an associated set of interaction techniques, this thesis describes the lessons we learned in the process of evaluating and improving our approach. We discuss how to package complex algorithms in a way that makes them accessible and practical; how to strike a balance between power and flexibility on the one hand, and simplicity and ease of use on the other; and how to address the unique set of challenges and opportunities presented when using computer vision in real-world applications. |
