Research On Key Technologies Of 3D Object Reconstruction Based On UV Representation

Three-dimensional(3D)reconstruction is an important issue in the field of computer vision.Its purpose is to use computer science technology to recover the 3D shape of an object from a two-dimensional(2D)image.In the fields of film,animation,games,medical imaging,human-computer interaction,and virtual reality,the demand for 3D object models has become normal,which also makes 3D object reconstruction a hot issue for scholars in recent years.Because different types of objects often have different characteristics in their 3D structures,the field of 3D reconstruction has gradually derived some branches of 3D reconstruction for specific types of objects.The 3D reconstructions of the human face and the human hand are the most representative tasks.This paper will conduct research on these two specific tasks.The UV-map-based 3D reconstruction is a classic 3D obj ect reconstruction method.It can efficiently achieve 3D reconstruction with a fully convolutional neural network by representing 3D objects as 2D UV maps.The methods based on the UV map do not rely on a priori parametric model and are easy to migrate to the 3D reconstruction task of other object models,and thus have a strong generalization ability.In this paper,the methods of 3D face reconstruction and 3D human hand reconstruction are researched based on UV mapping.At present,most methods of 3D face reconstruction and 3D hand reconstruction can achieve good results under controlled conditions,but their performance under unconstrained conditions is not ideal.Specifically,the lack of information caused by occlusion makes it difficult for the reconstruction network to infer the shape of the model in the unknown area;the large-scale change of pose will significantly increase the solution space of the problem and increase the modeling difficulty.According to the characteristics of the face model and the human hand model,this paper improves the robustness of the method under unconstrained conditions to achieve a better overall performance of the method.The contributions of this article are as follows:? Aiming at the problem of occlusion and large pose change in face reconstruction,this paper proposes to decouple the pose and shape of the face model to improve the robustness of face reconstruction.Aiming at the problem of large pose variace,this paper proposes a dual face representation method,that is,pose-dependent face and pose-independent face,which are represented as UV maps,and the pose is estimated by the alignment of them.The pose information is encoded in the pose-dependent face,and the pose obtained from this dense representation is not sensitive to the outliers and has stability;the shape information is encoded in the pose independent face,which has smaller solution space and is easier to learn.To solve the occlusion problem,this paper proposes an occlusion-aware attention mechanism,which highlights the effective features of the visible area of the face while preserving the context information.In addition,in the alignment process of face pose estimation,combined with the occlusion detection results,the more accurate nonoccluded vertices are aligned to further improve the accuracy of pose estimation.This work is evaluated on the AFLW2000-3D database and the Florence database.Our method is superior to the existing methods in face alignment and 3D reconstruction.? In order to solve the problems of large joint amount,large posture change,and frequent occlusion in 3D hand reconstruction,this paper proposes to model the uncertainty of vertex position by transforming the direct vertex coordinate regression into the confidence estimation based on the heat map,so as to improve the reconstruction accuracy.In this paper,the 3D heat map is decomposed into three 1D heat maps for regression,so as to reduce the space complexity and make it possible to be applied in the dense hand model.In this paper,we propose to establish a branch network at the lower level of the network to regress the 2D hand joints' heat maps,which can help the network to learn the pose information and improve the robustness to the pose variation.In addition,the joint regression branch is combined with the occlusion-aware attention mechanism to enhance the low-level features and improve the robustness to occlusion.In this paper,we evaluate our work on the FreiHAND database,and the evaluation results are better than those of the existing state-of-the-art methods.The ablation study of the proposed modules verifies their effectiveness.