Line Drawing Interpretation Using Belief Propagation
Yansheng Ming, Hongdong Li, Jun Jie Sun · 2011
The interpretation of line drawings of trihedral planer objects is a classic problem. In this paper, it is formulated as a Bayesian inference problem. Given a line drawing image, a Markov random field can be built whose nodes represent the labels of edges. Its clique potential functions are designed to encode the valid junctions in the Huffman-Clowes catalogue. The belief propagation algorithm is used to find the most probable labeling of the edges. We find this algorithm closely related to the arc consistency methods. However our probabilistic formulation can accommodate uncertainty in junction detection and make use of various image cues. These advantages are demonstrated in the experiments.