Multi-view and multivariate gaussian descriptor for 3D object retrieval

3D object retrieval is a hot research topic in computer vision domain, and several feature descriptors have been proposed, such as Zernike moments and HOG. However, multi-view images factor often be ignored in the feature extraction. Inspired by the Multivariate Gaussian descriptor and multi-view latent relationships, we propose a new feature descriptor called Multi-view and Multivariate Gaussian (MMG) Descriptor for 3D object retrieval. In detail, the local statistics of an image is characterized by using multivariate Gaussian distribution which is continuous and can effectively estimate different orders statistics in the local neighborhood. Furthermore, images from different perspectives are explored when extracting the characteristics of an object. Extensive experimental results on ETH dataset and 3Ddataset show that: 1) MMG descriptor is more suitable for 3D object retrieval than Zernike Moments and HOG whose performance is much better than that of other two descriptors; 2) The performance can also obtain some improvements when multi-view factor is considered. 3) When the different angles and number of images are chosen, their performances also have fluctuations.