Image-Based Triangular Surface Model Refinement

Image Matching based three-dimensional reconstruction is an important part of computer vision and photogrammetry.It mainly includes the steps of camera calibration,dense point cloud generation,mesh construction or mesh construction after point cloud fusion,and texture mapping,and is the most popular image based three dimensional reconstruction method nowadays.However,because of the assumption that the surface to be reconstructed is smooth and continuous,the influence of perspective projection deformation and noises,the surface constructed with image matching is usually rough and suffers loss of details.Variational surface evolution method regards the surface to be reconstructed as a continuous unknown function in three-dimensional space and defines a photometric consistency cost function on it.The surface model reconstruction problem is regarded as the minimum value solution of this cost function,which is usually solved by gradient descent method.Variatinal surface evolution can reconstruct high level surface details and depress the noises by using a regularization term.However,the evolution process is much too slow when the initial surface given is far from the real surface.The main content of this paper is image based three dimensional reconstruction by integrating image matching and variational surface evolution.Considering the efficiency and precision,we use semi-global matching to generate point cloud and then use s-t cut to construct an initial model.Finally,we use variational surface evolution to refine the model.The main contributions of this paper include three aspects: First,in view of the problem that the gradient descent based surface evolution method used by Vu et al.does not always converge when the depth differs greatly in the same camera view,we propose an image resolution based pixel-wise weighting method to balance the gradient of the surface according to the camera imaging model.Second,In order to solve the problem of surface roughness in the reconstruction result of weak texture planar region on the surface,a local window plane constraint on the depth map is used to restrain this phenomenon,which greatly improves the planar smoothness of the reconstructed surface.Finally,to speed up the surface refinement step,we use both CPU and GPU to accelerate the evolution progress,and the evolution efficiency is improved several times,which is of great significance to the practical application of this method.