多视密集匹配并行传播GPU-PatchMatch算法

针对多视密集匹配的效率较低的问题,提出了GPU-PatchMatch多视密集匹配算法。该算法使用GPU提高PatchMatch的计算效率;同时充分利用稀疏场景信息,对深度信息进行规则初始化;为提高传播效率,使用了金字塔红黑板并行传播深度信息。最后在DTU、Strecha和Vaihigen数据集上进行了试验,并与常用的多视密集匹配算法进行对比。试验结果表明,本文算法在重建效率上有较大提高,与CPU算法(PMVS、MVE、OpenMVS)相比有7倍以上提升,与GPU算法相比也有2.5倍以上提升,表明本文算法的有效性。

A GPU-PatchMatch multi-view dense matching algorithm based on parallel propagation

Aiming at the problem of low efficiency of multi-view dense matching, a GPU-PatchMatch multi-view dense matching algorithm is proposed. The algorithm uses GPU to improve the computational efficiency of PatchMatch. At the same time, it also makes full use of sparse scene information to initialize the depth information. In addition, in order to improve the propagation efficiency, it uses the pyramid red-blackboard to propagate the depth information in parallel. Finally, the experiments are carried out on the DTU, Strecha and Vaihigen datasets, and compared with the commonly used multi-view dense matching algorithms. The results show that our algorithm has a significant improvement in reconstruction efficiency, which is 7 times higher than the CPU algorithm (such as PMVS, MVE and OpenMVS), and 2.5 times higher than the GPU algorithm (such as Gipuma), which proves the effectiveness of the proposed method.