Shape-to-Shape Translation And High-Accuracy Reconstruction For Point Cloud

In recent years,the field of virtual reality and augmented reality has developed rapidly,and the concept of metaverse has also received widespread attention,which has led to a rapid increase in the demand for modeling 3D shapes in related fields.As the generation of two-dimensional images driven by artificial intelligence is becoming more and more popular,how to use artificial intelligence technology to realize content generation on three-dimensional shapes has attracted extensive attention from researchers in the field of computer graphics.In the shape generation task,since many application scenarios need to use a set of shapes with the same characteristics,such as furniture of the same style,how to batch generate shapes that meet specific requirements has become a problem worth studying.The point cloud shape-to-shape translation task is designed to address this problem.The task aims to decouple the encoding of point clouds into two parts,style and content,and then generate point cloud shapes of specified styles.But existing shape translation methods have some limitations.First,it is difficult for existing shape translation methods to disentangle shape information correctly;second,existing shape translation work is designed for translation tasks between two styles,and it is difficult to be directly applied to multi-style translation;third,existing work The translation effect of is subject to the reconstruction effect,but the autoencoder reconstruction network of point clouds often loses details,resulting in the translated shape not being similar to the original shape.To solve the first two problems,this paper proposes a new deep learning method,which aims to achieve shape translation from the unpaired shape domain in an unsupervised manner,and named the method USTNet.The core of the USTNet method is to use disentangled representation learning to distinguish the content and style features of 3D shapes,to obtain content and style encoding.Specifically,given two shapes from different style domains as the input of the network,USTNet decomposes their encoding into a style encoding that includes the differences between the domains and a content encoding that includes the common features between the domains.By fusing the style and content codes of the target domain shape and the original domain shape,USTNet can generate new shapes that are similar in style to the target domain and preserve the content features of the original domain shape.Since the style encoding space is shared among styles,USTNet can obtain shape interpolation between different styles by modifying the encoding in the space containing style information of different shape domains.Furthermore,the method can be directly used to achieve multi-domain translation of 3D shapes by changing the setting of modules in the network from binary classification to multi-classification.Experimental results show that USTNet can generate more realistic and natural translated shapes compared with 3DSNet.At the same time,the quantitative evaluation results prove that compared with other methods,USTNet has achieved better evaluation results in the quantitative evaluation.To solve the third problem,inspired by the work of Deep SDF,this paper proposes a method to bypass the performance bottleneck of the point cloud encoder,using the point cloud autodecoder structure which only contains the decoder to realize the point cloud.Cloud reconstruction,thereby improving the effect of point cloud shape translation and editing tasks.The autodecoder structure does not contain an encoder,but only a decoder,so the method is not affected by the performance of existing point cloud encoders.At the same time,the decoder in the autodecoder is not limited to a specific structure and can use a variety of mainstream decoder structures.In this paper,an autodecoder network based on three different decoder structures is designed,and the quantitative and qualitative experimental results prove that the point cloud autodecoder has obvious advantages over previous reconstruction methods,upsampling,and unsupervised representation learning tasks.It also achieves comparable results to existing methods on generation tasks.