Prior depth-based multi-view stereo network for online 3D model reconstruction

Abstract

This study addresses the online multi-view stereo (MVS) problem when reconstructing precise 3D models in real time. To solve this problem, most previous studies adopted a motion stereo approach that sequentially estimates depth maps from multiple localized images captured in a local time window. To compute the depth maps quickly, the motion stereo methods process down-sampled images or use a simplified algorithm for cost volume regularization; therefore, they generally produce reconstructed 3D models that are inaccurate. In this paper, we propose a novel online MVS method that accurately reconstructs high-resolution 3D models. This method infers prior depth information based on sequentially estimated depths and leverages it to estimate depth maps more precisely. The method constructs a cost volume by using the prior-depth-based visibility information and then fuses the prior depths into the cost volume. This approach significantly improves the stereo matching performance and completeness of the estimated depths. Extensive experiments showed that the proposed method outperforms other state-of-the-art MVS and motion stereo methods. In particular, it significantly improves the completeness of 3D models.