Improving spectral segmentation of 3D meshes using face patches

Abstract

A huge amount of research work has been devoted in recent years to segmentation of 3D meshes composed of planar triangular faces. In particular, spectral segmentation has had a fair share of this work because it is extremely faster than other segmentation techniques, especially those based on AI and machine learning. However, existing spectral segmentation techniques suffer from complex processing and heavy computation due to dealing directly with these faces. The present article is an attempt to address this issue by proposing an effective technique based on grouping the faces skillfully into higher-level structures called patches. Specifically, each patch is made of two neighbor faces, effectively cutting the number of low level structures processed by the segmentation technique into almost half. However, since the constituent mesh structures have changed from face to patch, the normal spectral segmentation methodology is altered to suit the new geometry. This alteration is reflected on the number of elements of both the eigenvectors and weight matrix, both reduced almost by 50\%. We have validated the proposed technique by segmenting numerous 3D meshes from public repositories. The resulting segments are colored in order to distinguish visually between different parts of the same 3D mesh. The experimental results indicate, both visually and quantitatively, that the proposed technique matches the performance of the best state-of-the-art methodologies, but at about half the time and space cost.