A Multi-View Stereo Evaluation for Fine Object Reconstruction

Current stereo matching methods based on end to end learning frameworks have shown strong results in the field of depth estimation, bringing significant improvements in robustness as well as flexibility in the accuracy and evaluation time trade-off. In this line of research we observe that the two sub-fields of binocular, and multi-view stereo have converged and are based on fundamentally the same architectures. In this work we aim to perform an objective comparison of these methods, controlling for architecture, and accounting for the rectification process typically used in binocular stereo. To our knowledge there is no prior work directly comparing the two. We aim to measure the performance of matching between rectified pairs, and plane-sweep based multi-view stereo. We test a range of camera configurations and studying the effectiveness of additional cameras in the context of a synthetic multi-view stereo dataset developed for evaluating 3D reconstruction in agriculture.