Research On Recurrent Attention Model Based 3D Shape Classification

Classification of 3D shapes is a fundamental problem in computer graphics and computer vision,and is significant to various applications such as traditional computer aided design and medical imaging as well as the mixed reality and robot navigation.In this thesis a novel method of 3D shape classification based on recurrent attention model is presented,separately using multi-view volumetric representation and multi-view projection images.And this method has achieved fine classification results.The method of 3D shape classification based on recurrent attention neural network is depths reinforce learning method in essence.It classifies 3D shapes as a sequence of decision making process,guided by the target and interacted with the environment.In each time step,the local observation of three-dimensional objects,according to the decision of the next action observed in the local region information and the current state of the environment,through the implementation of action effect observation of the environment of the system,the final integration of local information,get the 3D shape category labels.The main contribution of this thesis includes the following two aspects.1.A 3D shape classification method based on voxel representation and recursive attention model is proposed.Firstly,transform 3D shapes to voxel representation so that it is capable of convolution feature extraction,3D shape information and keep relatively complete.Secondly use the model at each step based on the current environment state prediction of three-dimensional space position next to observe the recurrent neural network based on visual attention,then the local area to the location of the center of the 3D convolution feature extraction,and feature extraction according to the observation of the position and update the current state of the environment.Thirdly,after a specified time step,according to the classification of three-dimensional objects in 3D space using historical interaction process represented by the state of the environment.2.Propose a 3D shape classification method based on multi view projection image and recurrent attention model.A new projection sampling method is proposed in this thesis.The 3D shape is projected from several views,which can cover the shape surface and avoid the excess information simultaneously.The deep residual network is used to compute the feature vectors of the projection images.Finally,the feature vectors of the projection images under different view are input into the recurrent attention neural network to classify the 3D shapes.This method simulates the way human beings observe three-dimensional objects,and chooses the spatial position of the next step according to the current environment.Experimental results on ModelNetlO and ModelNet40 datasets show that compared with existing methods,the proposed method achieves the leading or comparable accuracy of 3D model classification.Since the proposed method requires only a few steps,the accurate classification results can be obtained,so the efficiency of classification is obviously improved compared with the existing methods.