Deep Learning Based Shape Completion Technology

Three-dimensional data model,as a type of digital representation which is very close to the real world,has been booming in many application fields,such as model production,robotics,urban construction,autonomous driving,virtual reality and so on.With the boosting of these applications,the integrity,authenticity,and reliability of three-dimensional data models are urgently demanded.In the proceed of object modeling,the situations such as occlusion and jitter are inevitable,which results in the loss of information in the 3D data model and the incompleteness of the model structure.Thus,the missing parts of the model need to be predicted and filled.Most of the traditional model optimization methods are based on the geometry of the small hole filling,but they cannot repair the large model missing area.In order to complete the model with a large number of missing data,this paper discusses two methods to deal with the incomplete 3D model.One is to complement the model based on the Generative Adversarial networks,focusing on the completing structure of the prediction model and assimilating the predicted model with the input model.The other is the octree-based high-resolution model completion network,focusing on the model structure and details of the prediction fill,and then make the output model has more abundant information.As for the first method,considering the structural repair of the incomplete model,this paper uses a generative adversarial mechanism combined with an auto-encoder to endow the network a stronger generation ability.Specifically,on the one hand,the encoding-decoding structure is used to calculate the reconstruction error,thus ensuring the consistency of input and output,and at the same time,the generation-adversarial mechanism makes the output model more real and natural;on the other hand,the network,on the basis of improving the stability,can predict the complete structure of the incomplete model,so that the output model has a better visual result.Instead,the second method,consider the detailed completion of the incomplete model.It not only needs to repair the structure of the incomplete model but also needs to predict the model more precisely.Hence,the octree structure is used in this paper to efficiently store the feature information of the 3D model,thereby reducing the memory requirements of the model and reducing the amount of computation cost.In addition,the octree prediction network can predict more unknown structure information by deepening the number of tree layers,which is beneficial to the completion and repair of local details of the three-dimensional model.