The 3D mesh representation is widely used to model and visualize 3D objects,
especially those for which there is no known way to model using geometric
functions. An important problem that arises in the analysis of surface meshes
is that of mesh segmentation. Segmentation is applied in applications
such as feature detection and model fitting.
This paper presents a mesh segmentation algorithm that is based on
differential geometry and geodesic information. First the algorithm computes
the Gaussian and mean curvature at each vertex in order to classify the
surface type (peak, pit, minimal surface, or flat) of each vertex. Then
based on this classification, the algorithm grows segments, each of which is
labeled as being of one of the surface types. Geodesic distance is
computed using an efficient local approximate method rather than having to
store compute pairwise distances over the whole mesh as is the case with
an all-pairs shortest path running of Dijkstra's algorithm.
The segmentation resulting from the presented algorithm seems meaningful and
useful for protein docking applications. It is hard to tell from the paper
how useful this segmentation is for other applications. The authors claim it
is ``satisfactory". The proteins with less erratic variations in curvature
seem to have nicer segment shapes and boundaries.