New Composite Shape and Texture Descriptors for 3D Model Retrieval

Nowadays, problem of shape and texture for 3D retrieval is still a challenge research. Although several methods exist, but we still have a space to improve the performance. In this paper, we aim to improve our previous 3D shape features and inserting texture features. We first do pose normalization as a process of adjusting the size, location, and orientation of a given object in a canonical space and generate three types of non-color image rendered such as silhouette image, depth buffer image and contour image and one binary voxel. We get shape features from four independent Fourier spectra with periphery enhancement, which called multi Fourier spectral descriptor. We second generate both color image rendered and color voxel. We build a color histogram as texture features by using both color images rendered and color voxel based on distance and color level. Shape and texture features are finally combined together by linear summation. We conduct experiments based on the SHREC 2013 and 2014 track retrieval on textured 3D models. The experiment results show on how our method outperform in NN while using dataset SHREC 2013 and in FT and ST while using SHREC 2014. 