Synchronized spectral CNN for 3D SHape Segmentation

Shihav presentation

Presentation can be downloaded here.

In this presenation Shihav presented a different approach compared to the multiview rendering and volumentric representations. The spectral CNN approach (paper) is more focused on local represenation of points in a point cloud or a mesh. The method requires the graph laplacian and works by doing convolution as multiplication in the spectral domain. The in turn also yields a canonical representation of the 3D objects using functional maps. The functional maps paper will most likely be discussed in a future journal club meeting. The spectral decomposition of the mesh gives a good representation for working with clearly defined regions in a mesh, e.g. for segmentation like the authors present in this paper.