Higher Order CRF for Surface Reconstruction from Multi-view Data Sets

Abstract

We propose a novel method based on higher order Conditional Random Field (CRF) for reconstructing surface models from multi-view data sets. This method is automatic and robust to inevitable scanning noise and registration errors involved in the stages of data acquisition and registration. By incorporating the information within the input data sets into the energy function more sufficiently than existing methods, it more effectively captures spatial relations between 3D points, making the reconstructed surface both topologically and geometrically consistent with the data sources. We employ the state-of-the-art belief propagation algorithm to infer this higher order CRF while utilizing the sparseness of the CRF labeling to reduce the computational complexity. Experiments show that the proposed approach provides improved surface reconstruction.