Sparse prior guided deep multi-view stereo

Recently, the learning-based multi-view stereo (MVS) methods have achieved notable progress. Despite this, the cost volumes of those methods are not robust enough for depth inference with the existence of noise, outlier and occlusion. In this work, we aim to reduce the matching ambiguities and increase the discrimination of the cost volume. For this purpose, we first propose a sparse prior guidance strategy which incorporates the sparse points of structure-from-motion (SfM) into cost volume construction. Further, we introduce a geometry-aware regularization module to enhance the representative power of cost volume regularization, which could adaptively fit the local geometric shapes. The two modules are straightforward yet effective, resulting in robust and accurate depth estimation. The detailed experiments are conducted on two public datasets (DTU and Tanks & Temples). The experimental results show that with the two components, the representative methods ( MVSNet and CasMVSNet) could make a significant gain in reconstruction quality.