A landmark-free approach for automatic, dense and robust correspondence of 3D faces

Highlights：

• We construct an automatic algorithm using high-entropy points instead of landmarks, and take both global and local features of faces into consideration for accurate semantic correspondence.

• We consider topological correspondence and propose a mesh correction algorithm to filter out non-uniform local deformations, and this helps the construction of a compact 3D face model.

• We leverage a prior model with shape and normal statistics to handle hard samples with outliers, noises and expressions. This benefits robustness of the correspondence process with supervised domain knowledge of faces.