Heterogeneous Sensor Data Fusion: How Many Cameras are Needed for an Accurate 3D Reconstruction of Indoor Scene?

Vision systems, consisting of conventional color cameras and depth cameras have proved capable of solving the problem of indoor scene reconstruction with sufficient detail and satisfactory accuracy. However, there are few guidelines for how many cameras of each type are really needed to obtain a reconstruction with sufficient accuracy. At the same time, there has been little formal, quantitative, evaluation of the 3D reconstruction result, instead, comparisons are typically made subjectively. In the present work, we attempt to provide a feasible way to discuss the relationship between the number of cameras used and the final reconstruction accuracy. We present a general pipeline, from pre-calibration of the system to the final 3D point cloud output to demonstrate a full 3D reconstruction process for large static scenes, using a number of different system configurations. We evaluate and compare the reconstruction accuracy of these systems for a common target object in the scene. Our hope is that these results can offer some guidelines for camera selection based on the actual requirements for reconstruction accuracy.