Research On The Method Of 3D Face Reconstruction From A Single Image

3D face reconstruction from a single 2D face image is an important task in computer graphics and vision communities,which can facilitate many applications such as face recognition,face frontalization and facial animation.Recent methods typically learn a CNN-based 3D face model that regresses the coefficients of a 3D Morphable Model(3DMM)from a single 2D face image,then performing 3D face reconstruction or dense face alignment.However,most of the methods only consider one-to-one face reconstruction,ignoring the correlation or difference within the input images,which may lead to face identity change for a given face image with various facial poses.Additionally,the shortage of 3D annotated training data considerably limits performance of those methods,making the model can hardly deal well with the in-the-wild examples.Based on this major observation,we explore the technology of 3D face reconstruction from a single image in two aspects,i.e.,network structure construction and data augmentation.The contributions of our work can be summarized as follows.1.The thesis proposes a robust solution to regress the 3D Morphable face Model(3DMM)coefficients by carefully designing a novel “Siamese” Convolutional Neural Network(SCNN).Specifically,our model can learn a globally coherent non-linear function that maps the data evenly to the output manifold,and then imposing constraints to make the recovered 3D shapes robust to pose variants of the same identity,and meanwhile discriminative to different subjects.Therefore,the model can preserve the identity information for a given face image to assist face recognition.Experiments on the challenging database AFLW2000-3D and 300W-LP have shown the effectiveness of our method by comparing with state of the arts.2.In view of the fact that recent methods typically aim to learn a CNN-based 3D face model that regresses coefficients of 3D Morphable Model(3DMM)from 2D images which need abundant 2D-3D annotations training data.The thesis proposes a method that leverages 2D “in-the-wild” face images to effectively supervise and facilitate the 3D face model learning.Inspired by the idea that the Conditional Generative Adversarial Nets can be used for weak supervision without requiring groud truth.We concatenate the potential representation of the input images with their corresponding 3DMM coefficients,to avoid the requirement of sufficient 3D annotations for training,to improve of the accuracy and robustness of 3D face reconstruction.Experiments on multiple challenging datasets have shown that this method outperforms state-of-the-arts for 3D face reconstruction by a large margin.3.Different from 2D methods,3D face alignment aims to fit a 3DMM to a 2D face image,it can inherently provide the visibility of each model point without any additional estimation,making it possible to deal with the self-occluded and large poses of face.The thesis proposes a method in which the dense 3D face model is fitted to the image and combined with the large pose face dataset(300W-LP)for model training.Experiments on the challenging AFLW2000-3D database have shown that this approach achieves significant improvements over state-of-the-art methods,especially on large poses(from 60° to 90°).