Industry-relevant benchmark dataset for the evaluation of 3D reconstruction methods and evaluation of different NeRF methods

This study assesses the potential of different Neural Radiance Field (NeRF) and 3D Gaussian Splatting (3DGS) methods in terms of the rendered images and extracted point clouds against classical Multi View Stereo (MVS) with a focus on industrial objects. Firstly, a benchmark dataset is created and captured with an industrial robot arm and camera setup. The camera poses are derived in two alternative versions, which are compared with each other: with Structure from Motion (SfM) and with the robot, by utilizing the kinematics of the robot and the hand-eye pose. The robot-derived poses are more robust and independent of the captured scene. Moreover, they are metric and can be determined much faster than the SfM poses. Secondly, current NeRF and 3DGS methodsvare evaluated on the benchmark data set, each with both pose versions. The results show that with both pose versions, similarly accurate images from novel views can be rendered. However, the images rendered with SfM poses are less accurate for complex objects. The extracted point clouds of trained NeRF and 3DGS models are usually less accurate than standard MVS point clouds, except for some transparent objects