Research And Implementation Of 3d Body Reconstruction With Dressed-human Photos In Wild

3D body reconstruction has always been a hot research direction in computer graphics.It has been widely used in video games,films,animation and other application.But for ordinary individual users,these traditional methods of 3D body reconstruction require high acquisition equipment and complicated acquisition methods,which are full of inconveniences.This thesis proposes a set of end-to-end automated 3D human body reconstruction techniques based on deep learning methods.A 3D human body model is directly reconstructed from the photos of the dressed human body in wild.The algorithm is divided into three main steps.The first step is to extract the silhouettes of the dressed human body from the photos in wild using a convolutional neural network based on visual saliency.The second step is to regress to get the body 3D key points from the silhouettes of the human body using a another convolutional neural network,and finally to get the 3D body model parameters bound by the 3D key points.The main work of this paper is as follows:1.This thesis has construct a dataset of human photos in wild.This dataset has made a contribution of getting good marks of extracting the silhouette of the human body from the photos in wild.The image segmentation model trained by this dataset also has a good robustness and generalization.2.This thesis proposes and designs automatic segmentation networks to extract the silhouette of the human body from the photos in wild based on visual saliency.This network first extracts high-level aggregation features and low-level aggregation features through feature extraction networks.Pyramid upsampling construction,deep supervised strategy,residual optimized structure and mixed loss function are also applied in this network to help get the best result than other existing semantic segmentation networks under our test dataset.And this is also an end-to-end automation method.No user interaction is required.The user inputs the photo taken in wild,and it directly outputs the the silhouette of the human body with dressed clothes.3.This thesis proposes a 3D keypoint regression algorithm based on convolutional neural networks.Based on the dressing human database,use The convolutional neural network directly returns the coordinates of the 3D key points of the human body.Compared with the original methods which need to specify the initialization coordinates,this algorithm realizes the end-to-end automation effect.