Guest Editorial: Special Issue on Structured Prediction and Inference

Computer vision has been profoundly influenced by machine learning in the past two decades. Canonical papers in the field, such as [Turk and Pentland(1991)], have applied statistical methods to visual data to achieve results that are both compelling and accurate. Classification algorithms, such as support vector machines, are now commonly applied to discriminatively train vision systems. Such systems empirically minimize risk functionals that measure the expected classification error rate given an i.i.d. sampling assumption [Vapnik(1998), Scholkopf and Smola(2002)]. While many intermediate goals in computer vision can be formulated as classification, regression, or dimensionality reduction– the settings most commonly addressed in the machine learning literature–more appropriate statistical methods are needed to train vision systems that make collections of related predictions, such as in segmentation, parts based models, or scene layout analysis. In these settings, the independence assumptions made by binary classifiers no longer hold, and independent prediction may not be computationally or statistically feasible. Structured output prediction [Bakir et al(2007)Bakir, Hofmann, Scholkopf, Smola, Taskar, and Vishwanathan, Nowozin and Lampert(2011)] is the task of predicting related variables given some input, such as image or video data. Study of structured output prediction in the machine learning field has led to a number of techniques be-