Research On 3D Shape Correspondence And Segmentation

Recently,the development of 3D scanning technology has brought opportunities for the construction of complex 3D models.The proliferation of 3D models has brought new challenges to 3D shape analysis and understanding.Via Al algorithms,3D shape analysis is aimed to explore consistent structures,semantic relationships,and also to infer model functionality between 3D models,such as grasping.How to accurately describe and effectively realize 3D shape analysis is the basis for intelligent robots to understand the three-dimensional world and shape.With the development of deep learning theory and application,3D shape analysis has not only shifted from traditional manual design features to feature extraction methods based on deep learning,but also the scale of shape analysis has changed from single or small set of shapes analysis to joint analysis of large 3D dataset.At present,the research about feature extraction methods based on deep learning and large-scale three-dimensional shape analysis is one of of interested topics on computer graphics.In this paper,3D shape correspondence and segmentation problems are be solved based on the deep learning method.In addition,an interactive 3D shape segmentation algorithm is proposed for fine shape segmentation.The main contributions of this paper include:(1)For 3D shape correspondence,a learning framework of SiamesePointNet is proposed.The framework is used to extract the consistent semantic features between the 3D shapes,and a feature matching method is used to realize the dense correspondence and key points retrieval between 3D shapes.Specifically,in order to more accurately mine the semantic structure features between models.Firstly,a deep cloud network framework of Siamese network is proposed.Secondly,based on the idea of metric learning,a large number of 3D model training sets based on component correspondence and N-tuple contrastive loss are used.Finally,considering the problem of high degree of freedom and generalization of the point cloud convolutional network framework,the global feature contrast loss constraint between 3D models is proposed.Experimental results verify the effectiveness of global feature constraints.In addition,by adding feature transform modules related to global features,the semantic expression ability of point-by-point features between shapes is further improved.This method can effectively solve the dense correspondence and key points correspondence between 3D shapes with large deformation,noise and partial missing.(2)For the problem of complex 3D shape segmentation,considering the influence of boundary detection on segmentation results,a compact feature representation of 3D shape is learned,then the pre-segmentation and boundary detection of 3D shape are completed.Given the learned semantic boundary,a fine segmentation result can be obtained.Firstly,using the image convolution neural network,a feature extraction model is established to learn the compact features of the 3D models.For this network,a dropout technique is introduced to improve the generalization of the network in feature extraction,and to realize pre-segmentation and boundary detection of 3D shapes.Second,based on the detected boundaries,a continuous,smooth and closed semantic boundary distance field is constructed.Finally,the distance field is introduced into the Graph-Cut model to guide to achieve fine segmentation.The experimental results show that the optimization algorithm based on semantic boundary guidance can significantly improve the accuracy of shape segmentation,especially for complex shapes.(3)For more detailed and diverse segmentation tasks,the multiple random walkers(Multiple Random Walkers,MRW)method is proposed to interactive segmentation algorithm.The interactive walk strategy between multiple walkers(using the restart probability of multiple walk-ers on the graph model)ensures the full exchange of information on the model.The confrontation between different walkers is the key to ensuring the accuracy of the final segmentation results.At the same time,this paper also extends the strategy of multiple random walkers to the cosegmentation task on model set.Compared to the traditional random walk algorithm,the interactive segmentation based on MRW is more stable,and the accuracy of 3D shape segmentation and co-segmentation can be achieved.In addition,the algorithm proposed in this paper can also realize the multi-scale segmentation of 3D shapes according to the users interaction and complex 3D shapes segmentation task.