Modeling of Tree Branches by Bayesian Network Structure Inference

In the paper, we present an approach to inferring 3D subtree structures from image pairs. The 3D structure is treated as a hidden Bayesian network, of which each node corresponds to an attributed skeleton point. The network structure is inferred in a bottom-up fashion. At the beginning, the root node of a subtree is manually specified in the images and then computed using stereo triangulation. Next, the subsequent computation automatically infers the child nodes stage by stage along the branches. At each stage, the child node states are sampled from a posterior distribution, which incorporates image observations in different viewpoints and pre-defined priors, such as smoothness. A tracebased stereo matching algorithm is introduced to propose the child node candidate states for computation efficiency. The experiments demonstrate that the proposed approach is competent in subtree construction.