Feature Encoding of Spectral Signatures for 3D Non-Rigid Shape Retrieval

Abstract

As the Internet and 3D modelling tools have led to an increasingly growth in the number of available 3D models, it becomes necessary to have a proper and smaller representation for searching purposes that captures the most important information about shapes. A large number of encoding methods have been proposed in the literature to create shape signatures from local descriptors. Two encoding methods have been receiving most attention from researchers given its informative characteristics: Fisher Vector [9] and Super Vector [36]. We propose to use these encoding methods combined with spectral signatures to represent 3D shapes. Although spectral signatures have many desirable properties to describe 3D shapes, for instance being invariant under rigid transformations and stable against non-rigid transformations, they do not perform so well in recent benchmarks. We propose improvements to the Wave Kernel Signature by analysing its behaviour when combined to different encoding methods for the purpose of shape retrieval and classification. At the end, we show a comparison of our method in two recent benchmarks.