@TechReport{Franc-TR-2012-16, IS = { zkontrolovano 15 Jan 2013 }, UPDATE = { 2012-10-01 }, author = {Franc, Vojt{\v e}ch and Antoniuk, Konstantyn and U{\v r}i{\v c}{\'a}{\v r}, Michal}, title = {Discriminative structured output learning from partially annotated examples}, institution = {Center for Machine Perception, K13133 FEE Czech Technical University}, address = {Prague, Czech Republic}, year = {2012}, month = {July}, type = {Research Report}, number = {CTU--CMP--2012--16}, issn = {1213-2365}, pages = {18}, figures = {2}, authorship = {70-15-15}, psurl = {[Franc-TR-2012-16.pdf]}, project = {PERG04-GA-2008-239455 SEMISOL}, annote = {The discriminative structured output learning has been proved successful in solving many real-life applications. A big deficiency of existing algorithms like the Structured Output SVMs is the requirement of fully annotated training examples. In this report we formulate a problem of learning the structured output classifiers from partially annotated examples as an instance of the expected risk minimization. We show that the minimization of the expected risk is equivalent to the minimization of the partial loss which can be evaluated on partially annotated examples. We proposed an instance of the partial learning algorithm for the class of linear structured output classifiers which we call Partial-SO-SVM. The Partial-SO-SVM algorithm leads to a hard non-convex optimization problem. We provide an algorithm solving the Partial-SO-SVM problem approximately using an additional prior knowledge about the problem. We demonstrated effectiveness of the proposed method on two real life computer vision problems, namely, the face landmark detection and the image segmentation.}, keywords = {structured output learning, partially annotated data}, }