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

COMMENTS ON OTHERS:

            Jonathan  

SUMMARY

              The paper proposes recognition of hand-drawn diagram structures using Bayesian conditional random fields (BCRFs). BCRFs are a generalization of Bayes point machines or discriminative Bayesian classifiers for single elements. The method jointly analyzes all drawing elements in order to incorporate contextual cues. According to authors, the advantage of BCRFs over conventionally-trained CRFs include, model averaging and automatic hyper parameter tuning.

            The application of BCRFs to ink classification is to discriminate between the containers and connectors in a drawing of organization charts. This particular task includes, subdivision of pen strokes into fragments, construction of conditional random field on fragments, BCRF training and inference on network. In the first step, the strokes are divided into simpler components called fragments. Fragments should be small enough to belong to a single container or connector. In contrast, strokes occasionally span more than one part when drawn without lifting a pen. The fragments are selected as strokes that form a straight line segments. In the process of conditional random field on the fragments, each fragment is represented by a node in the network. 

DISCUSSION

            The proposed work shows an interesting method to segment ink features from a sketch application. The method is mathematically rigorous and therefore hard to understand by the content. Great deal of information of the feature extraction omitted from the paper, and therefore makes the works shine a bit less. Also note that the BCRFs accuracy is less compared to the joint performance of BCRFs-ARD.