3D Object Recognition And Retrieval Based On Multi-View Feature Fusion

3D object recognition is a fundamental task of 3D data analysis,which is becoming an important reach area of computer vision.With the development of deep learning,convolution neural networks(CNNs)have shown encouraging results in 2D image processing.The multiview method is one of the mainstream methods for analyzing and processing 3D objects.This method uses a CNN to analyze the projections of different perspectives of 3D objects,and infers the global information of objects by fusing the features of multiple views.This paper focuses on the key techniques of multi-view methods in 3D data analysis.In this paper,the view-point selection,feature extraction and feature fusion modules of the multi-view method are deeply investigated,and improved models and algorithms are proposed.For view-points selection,a new strategy is presented.Our strategy proposes to place viewpoints evenly on an octahedron circumscribing a 3D shape,under which we can control the intensity of generated view-points by setting different density levels.By establishing an undirected graph structure between viewpoint points and clustering into multiple groups according to spatial relationships,hierarchical feature fusion operations are realized.For feature extraction,a view feature mapping module is proposed.The module learns a mapping matrix by a multi-layer perceptron to transform adjacent features to a new feature space.Moreover,we modified the loss function to constrain that neighboring features stay close to each other,guiding the network to learn refined local features and structural information of objects.For view feature aggregation,we propose two different approaches base on convolution.The first one aims at aggregating feature vectors from the CNN.Neighboring view features are together according to their adjacent relations and consecutive convolution operators are applied on them for further feature abstraction.The second one is directed to the feature map containing the spatial structure information,and the maximum response is obtained by using the max-pooling to capture the correlation between the adjacent views.Through the above two convolutions,the effective fusion of multi-view features is realized,and the global information of the object is obtained.In the experimental part,the performance of the algorithm is evaluated on the standard dataset commonly used in 3D object recognition,and compared with other state-of-art approaches.The effectiveness of the proposed algorithm is demonstrated by experiments and analysis.