Reading #19: Diagram Structure Recognition by Bayesian Conditional Random Fields (Qi)

Comments

Sam

Summary

This is a top down recognizer that relies heavily on context to determine the correct classification of each stroke. In this case a model of Bayesian Conditional Random Fields is used to determine the classification of the strokes. Each stroke that is classified affects the classification of it neighbors. The paper provides a deep mathematical background towards the model compared to others. The first step in recognition is to fragment the strokes in order to create the Bayesian CRF. Note that the fragment here is defined in a different way than in other papers. It is not the line formed by each 2 consecutive point in the strokes, but the set of points in the stroke that could be recognized as a straight line. This implies corner detection as seen in previous posts. Then they can construct the BCRF and train it to make inference on the network. The results show different classifications on variations of the CRF, showing that the BCRF behaves better in the recognition. An improvement of Automatic Relevance Determination makes the recognition even better.

Discussion

A nice thing about this work is that it takes a concept of another field, (computer vision) and applies it succesfully to the domain of sketch recognition. It is not the first time that we see this phenomena. Since the sketch recognition is such an open field in the moment many works successfully or not attempt to convert the sketch recongnition problem into a more familiar one (fuzzy logic, graph searches, HMMs…). In this case, the Bayesian Conditional Random Fields show interesting results in this domain.