Multi-distribution fitting for multi-view stereo

We propose a multi-view stereo network based on multi-distribution fitting (MDF-Net), which achieves high-resolution depth map prediction with low memory and high efficiency. This method adopts a four-stage cascade structure, which mainly has the following three contributions. First, view cost regularization is proposed to weaken the influence of matching noise on building the cost volume. Second, it is suggested to adaptively calculate the depth refinement interval using multi-distribution fitting (MDF). Gaussian distribution fitting is used to refine and correct depth within a large interval, and then Laplace distribution fitting is used to accurately estimate depth within a small interval. Third, the lightweight image super-resolution network is applied to upsample the depth map in the fourth stage to reduce running time and memory requirements. The experimental results on the DTU dataset indicate that MDF-Net has achieved the most advanced results. It has the lowest memory consumption and running time among the high-resolution reconstruction methods, requiring only approximately 4.29G memory for predicting a depth map with the resolution of 1600 x 1184. In addition, we validate the generalization ability on Tanks and Temples dataset, achieving very competitive performance. The code has been released at https://github.com/zongh5a/MDF-Net.