Deep Learning Based 3D Shape Saliency Detection,Classification And Segmentation

In computer graphics and related areas,3D representations of objects(e.g.,mesh and point cloud)is the foundation of understanding and manipulating objects.By analyzing shapes of 3D objects,machines are able to imitate the human being's intelligent behavior,which is a promising technology in intelligent robot and automatic driving.Nowadays,deep neural networks are widely used to extract information from low-level features of objects and learn high-level semantic features for a specific task.However,the irregular spatial structures in 3D data lead to a big challenge for applications of deep learning.Recently,deep learning community began to pay attention to 3D vision tasks,and some novel network frameworks were proposed.In the areas of 3D vision and shape analysis,a lot of problems remain to be exploited and solved effectively.In this paper,we attempt to apply deep learning technology to two problems:· For mesh representation of shapes,the mesh saliency detection problem is studied.We develop a salient region detection method based on the feature-fusion learning,which is suitable for different kinds of shapes.Starting with computing several features of the shape,the low-level features are organized in a multi-scale way and input into a 1D convolution neural network.After optimizing a central regularized loss function,we get the high-level and discriminative feature,and the prediction of salient regions.· For point cloud representation of shapes,the method for point cloud classification and segmentation is developed.We propose a new point cloud network with a hybrid architecture to solve these tasks.Taking consideration of geometric and topological structures of shapes,the network makes use of both two kinds of information to learn parameters,leading to promising performance.