Abstract
Matching cost aggregation plays a fundamental role in learning-basedmulti-view stereo networks. However, directly aggregating adjacent costs canlead to suboptimal results due to local geometric inconsistency. Relatedmethods either seek selective aggregation or improve aggregated depth in the 2Dspace, both are unable to handle geometric inconsistency in the cost volumeeffectively. In this paper, we propose GoMVS to aggregate geometricallyconsistent costs, yielding better utilization of adjacent geometries. Morespecifically, we correspond and propagate adjacent costs to the reference pixelby leveraging the local geometric smoothness in conjunction with surfacenormals. We achieve this by the geometric consistent propagation (GCP) module.It computes the correspondence from the adjacent depth hypothesis space to thereference depth space using surface normals, then uses the correspondence topropagate adjacent costs to the reference geometry, followed by a convolutionfor aggregation. Our method achieves new state-of-the-art performance on DTU,Tanks & Temple, and ETH3D datasets. Notably, our method ranks 1st on the Tanks& Temple Advanced benchmark.