Shape completion using orthogonal views through a multi-input–output network

Knowing the shape of objects is essential to many robotics tasks. However, this is not always feasible. Recent approaches based on point clouds and voxel cubes have been proposed for shape completion from a single-depth view. However, they tend to be computationally expensive and require the tuning of many weights. This paper presents a novel architecture for shape completion based on six orthogonal views obtained from a point cloud (they can be seen as the six faces of a dice). Our network uses one branch for each orthogonal view as input–output and mixes them in the middle of the architecture. By using orthogonal views, the number of required parameters is significantly reduced. We also introduce a novel method to filter the output of networks based on orthogonal views and describe algorithms to convert an orthogonal view to voxel cube and point cloud. We compared our approach against state-of-the-art approaches on the YCB and ShapeNet datasets using the Chamfer distance and mean square error measures and showed very competitive performance with less than 5% of their parameters.