3D CNN based Partial 3D Shape Retrieval Focusing on Local Features

In this paper, we propose a new method for 3D CNN based partial 3D shape retrieval focusing on local features. A 3D partial shape in our approach is defined by a collection of points on the visible surface projected on the view-screen, during the rendering of a given 3D shape. We construct a voxel from the partial 3D points after extracting the local feature vectors and subsequent dimensional reduction by PCA (Principla Component Analysis) and feeding the reduced feature vectors to 3D CNN. This is a unique approach in contrast to the traditional approach to 3D CNN where the voxels have their values either Os or 1s (i.e. binary voxels). We conducted experiments with a SHREC2016 partial 3D dataset. Our proposed approach outperformed the VoxNet. We also compared our proposed method with other previous methods for partial 3D shape search.