A novel 3D dense reconstruction with high accuracy and completeness

This paper presents a planar patch based method for 3D dense reconstruction. The object surface is represented as a point cloud. For high reconstruction accuracy and completeness, our strategy is to fit the object surface by planar patches at different scales. Furthermore, we apply adaptive weights to define the photo-consistency measures for view matching. To optimize the view matching, we adopt a derivative-free GLN-PSO stochastic optimization method. Finally, to improve the reconstruction quality we design a patch priority queue for ordering the seed patches for patch expansion. The experimental results indicate the current implementation results are generally superior or comparable in comparison with the top ranked reconstruction methods reported in the Middlebury MVS website.