3D Model Retrieval Via Class-statistic Model And Group-pair Convolutional Neural Network

With the increasing application of 3D models,3D model retrieval algorithms have gradually become a hot research topic.At present,the exciting 3D model retrieval algorithm based on multi-view is to process the each view of the object as independent description information while ignoring the correlation among the views of the object.And the existing algorithms still have a lot of room for improvement in retrieval accuracy and speed.Therefore,based on these problems,the research work in this paper mainly includes the following parts: 1)A 3D model retrieval algorithm based on the class-statistic model and the triple-constraint function is proposed;2)A 3D model retrieval algorithm based on group-pair convolution neural networks is proposed;3)A 3D object recognition algorithm based on pairwise multi-view convolutional neural networks is proposed.Specific work as follows:1)A 3D model retrieval algorithm based on the class-statistic model and the triple-constraint function(CSTC).In this algorithm,a class-statistic model is first proposed.This model constructs Gaussian models that describe the distribution of the data in each dimension by analyzing the distribution of each dimension of all objects in each class.Then,the influence factors of the feature dimensions are calculated by analyzing the convergence degree of each dimension.Afterwards,the class-statistic model is used to calculate the probability that the object belongs to each category,and then,the category information of the object is obtained.In addition,in order to improve the retrieval speed,the measurement strategy based on the object distance is used instead of the measurement strategy based on the view distance.In order to improve the retrieval precision,a pairwise model is constructed to strengthen the restriction of the similarity between objects.Finally,the 3D model retrieval algorithm based on the class-statistical model and triple-constraint function is constructed based on the fusion of the probability result of the class-statistic model,the object distance information and the restriction of the pairwise model.The experimental results in ETH,NTU-60,MVRED and PSB 3D datasets show that this algorithm has a significant improvement in retrieval accuracy and speed.2)A group-pair convolution neural networks retrieval algorithm(GPCNN)is proposed.In this algorithm,an end-to-end double branch network structure is designed,and each branch can receive multiple views.At the end of the first part of the network structure,the view pooling layer is introduced to mine the correlation among the multiple views so as to enhance the contrast of objects on two branches.Due to the small number of objects in existing 3D multi-view datasets,the training of the network is not sufficient and it’s easy to occur the under-fitting.Therefore,in order to expand the number of training samples,the group-pair view is proposed for this network so as to directly increase the number of network training samples.Similarly,experimental results on three datasets ETH,NTU-60,and MVRED show that the multi-view information enhances the contrast between objects in different class and improves retrieval performance.On the other side,end-to-end Training can improve the overall training effect.3)In practical applications,3D object recognition is widely used.Therefore,the GPCNN network is extended to a pairwise multi-view convolutional neural networks(PMVCNN).The network structure still inherits the GPCNN end-to-end and double-branch characteristics.The difference between PMVCNN and GPCNN is that,after the view pooling layer,the features are merged into the third part of the new network structure,and the classification loss function is used as a training guide.The advantage of this algorithm is that not only the similarity between the two objects can be obtained according to the classification result,but also the category of the object can be deduced according to the compared object category to achieve the object recognition target.In order to better realize the object recognition process,an induced set is introduced in this algorithm.The induced set is a small part of data set and not be used in the training but only used as the guidance collection of object identification for testing.Experiments on ETH and NTU-60 datasets show that PMVCNN has good performance in object recognition and has a wide range of application prospect.