A cluster-based PMVS algorithm with geometric constraint

Patch-based Multi-view Stereo (PMVS) is a simple and effective algorithm for generating colorful 3D points with a set of pictures and camera parameters but without any initialization. Still, there are some problems when large image sets are reconstructed. On the one hand the reconstructed surface is not smooth, continuous enough, and the problem becomes more serious under certain image capturing configuration such as downward-shooting or upward-shooting. On the other hand, time and space complexity are too high to reconstruct. This paper proposes a novel algorithm for 3D reconstruction based on clustering. Firstly, decomposing the images into a set of overlapping subsets of pictures to reduce the running time and solve the problem that system memory consumption is too large, then reconstruct each cluster through PMVS algorithm with geometric constraint to improve the accuracy and smoothness, finally merge all resulting reconstructions. Experimental results show that the proposed method is available and practical.