Structured prediction models for online sketch recognition
Adrien Delaye · 2014
In this work we investigate the application of structured probabilistic models for the interpretation of online handwritten sketches. A exible and fully trainable approach is presented, comprising a method for training pairwise stroke distance and a structured prediction model for labeling strokes. Dierent variants of model structures are proposed for modeling the context. No symbol hypothesis generation is performed before stroke labeling and hard segmentation decisions are deferred to the post-processing stage. The system is validated by experiments showing high prediction accuracy over two dierent data sets of online handwritten sketches.