Automatic And Efficient Reconstruction Of Human Body Based On Camera Array

3D reconstruction based on images is a classical research subject in computer vision.There are many important applications in a lot of region such as Remote Sensing,Augmented Reality,Virtual Reality,Auto Drive and so on.In 3D reconstruction of human body,the camera array system can take photos in different views at the same time and get a good reconstruction result with the intrinsic parameters and extrinsic pararmeters.It is becoming more and more common in 3D reconstruction with GPU.When the images have the textureless area in 3D reconstruction,it is not good enough to get a good result.So we propose an approach with Visual Hull to constrain the depth and normal.It improves the result in textureless area.At the same time,we use a segmentation algorithm based on Fully Convolutional Network to segment the person and background to get the mask for the VisualHull.Base on the Schonberger's PatchMatch method,we propose an odd-even iteration view selection algorithm.We are inspired by the thought of Generalized Expectation-maximization algorithm including E-Step and M-Step.On the one hand,we use hidden markov model to select views as E-Step.On the other hand,we choose more neighbour's pixel information to recover the depth in PatchMatch algorithm as M-Step,and it converges fast.And cause the Schonberger's work did not parallel completely,we propose a new paral-lel method with odd-even itetration to get a pixelwise parallel result.Under the GTX 1080 Ti,we speed ups to twice as the Schonberger' s method.As the 3D reconstruction develops,automatic and efficient 3D reconstruction demand grows,we design a tool integrated human body segmentation,depth recovery,surface recon-struction and textures.People need to provide images and camera parameters,it can output a 3D model with a good textures.