When humans "recognize" things one of answers can always be "unknown" or "that's new." Existing vision and machine learning research has made great progress but have done so in a closed set paradigm - which explicitly minimizes risk/errors over what is known. As a computer vision system is moved toward real problems, it must face up to an open world. This project develops technologies for a new fundamental theory of "open vision" and corresponding set of tools that are explicitly designed to address open set recognition. At the heart of this research are three key concepts: 1) extending classical learning theory to include the risks of labeling open/unknown spaces, and then building classifiers that balance empirical risk, smoothness and open space risk; 2) meta-recognition - bringing a statistically-grounded probabilistic interpretation to classifiers, improving their ability to produce "confidence" in their answers; 3) operational adaptation - developing new approaches to address, at operation/run time, missing data or new data incorporating both open set and meta-recognition technologies. The work is also developing new approaches for open set evaluation, addressing problems in face recognition and visual object recognition as well as adapting classical machine learning datasets.
The open vision paradigm is embodied in open source tools that provide performance at or significantly advancing the state of the art while providing greater protection form unknown unknowns. Since most science is exploring the unknown, providing easy to use open source learning/recognition tools design for both known and unknown data, the project have broad impact to many different applications.
