3D Indoor Scene Classification using Tri-projection Voxel Splatting

Abstract

In this paper, we propose a new method for classifying a 3D scene. A 3D scene consists of a large collection of 3D shape objects. The complexity of a 3D scene makes it hard for us to classify which 3D scene we are dealing with. For instance, a 3D scene of a “room” may be an office, a kitchen, a living room, or a bed room. Here we propose a novel approach to classifying a 3D scene with Tri-projection Voxel Splatting (TVS), taking into account the voxel density along the depth direction. In TVS we first normalize a 3D scene in terms of position and size, followed by converting data into point clouds, via voxelization, and by projecting the scene on three perpendicular planes, reflecting the voxel density along the depth direction. Subsequently, we merge the three projected images, and we finally apply deep learning to predict the class of each 3D scene. To demonstrate the effectiveness of our proposed method (TVS), we conducted experiments with 3D indoor scene dataset extracted from Princeton University's SUNCG dataset. From the experiments, our proposed method outperformed the previous methods.