Structured Prediction Models for High-level Computer Vision Tasks

Опубликовано 17 августа 2016, 3:26

Rich statistical models have revolutionized computer vision research: graphical models and structured prediction in particular are now commonly used tools to address hard computer vision problems. I discuss what distinguishes these computer vision problems from other machine learning problems and how this poses unique challenges. I argue that most computer vision models are misspecified and discuss the consequences of popular estimators in this case, concluding that we either have to use non-parametric models or use estimators robust to misspecification. As one possible solution, I propose a novel discrete random field model applicable to a large number of computer vision tasks. The model is conditionally specified, non-parametric, and able to represent complex label interactions, yet it can be trained from hundreds of images in minutes on a single machine.