The objective of the project is to explicitly label the mesh into its components. The problem of mesh labelling can be viewed a classification problem as such this project trains a random forest classifier trained on mesh data generated using a ensemble of mesh descriptors which effectively and concisely capture the surface,volume and orientation of the mesh.
- Average Geodesic Distance (Local surface property)
Approximate average geodesic distance using fast marching algorithm.
- Shape diameter function(Local Volume property)
Approximation method calculate the distance from surface to the medial axis. Uses Moller-Trumbore algorithm for triangle ray intersection.
- Curvature (Gaussian,Mean and principle curvatures)(1-Ring neighbourhood surface property)
First order differential attributes on a piecewise linear triangular mesh.
- Volumetric shape images(Local Volume property)
Better approximation of distance from surface to the medial axis. A Two pass algorithm which internally uses SDF.
- Shape Context(Surface and mesh face orientation property)
Combines geodesic distance with the orientation by computing the angle between the surface normal and line connecting the face centres.
A Random Forest Classifier was able to label the mesh with an accuracy of 92% and an SVM Classifier with accuracy of 93%.
Time to generate each features
Please Checkout PresentationFolder for detailed report.