The paper proposes a novel method called GoMVS for multi-view stereo (MVS) reconstruction. The key idea is to aggregate geometrically consistent costs by leveraging local geometric smoothness and surface normals, which allows better utilization of adjacent geometries.
Specifically, the method first constructs a cost volume using multi-scale image features and differentiable homography. It then introduces a geometrically consistent aggregation scheme, which consists of two main components:
Geometrically Consistent Propagation (GCP) module: This module computes the correspondence from the adjacent depth hypothesis space to the reference depth space using surface normals, and then propagates the adjacent costs to the reference geometry.
Aggregation using convolution: After propagating the adjacent costs, a standard convolution layer is used to aggregate the geometrically consistent costs.
The authors also investigate different choices for obtaining surface normals, including using depth-computed normals, cost-computed normals, and off-the-shelf monocular normal estimation models. They find that the monocular normal estimation model performs well across different datasets.
Extensive experiments on the DTU, Tanks and Temples, and ETH3D datasets demonstrate that the proposed GoMVS method achieves new state-of-the-art performance, particularly in terms of completeness of the reconstructed point clouds. The authors attribute this to the ability of their method to better utilize adjacent geometries through the geometrically consistent aggregation scheme.
Til et andet sprog
fra kildeindhold
arxiv.org
Vigtigste indsigter udtrukket fra
by Jiang Wu,Rui... kl. arxiv.org 04-12-2024
https://arxiv.org/pdf/2404.07992.pdfDybere Forespørgsler