Weakly Supervised Part Segmentation And Anti-rotation Classification And Segmentation Of 3D Models Based On Sphere-node Graph

In recent years,3D data processing has gained increasing attention in the areas of computer vision,computer graphics,virtual reality and industrial design.3D data can be represented in a number of different ways,including point clouds,voxels and meshes.The sphere node graph is a novel 3D model representation proposed in the last two years,which contains richer local and global shape structure information than surface point clouds,and is more concise than representations such as voxels(or octrees).Deep neural networks have facilitated the processing of 3D data,and their performance in a wide range of tasks has vastly outperformed traditional methods.Semantic segmentation of 3D models and 3D model classification are two of these important tasks.Due to the rapid development of sensors and data processing techniques,3D datasets are becoming increasingly large,and supervised learning methods require extensive and exhaustive manual annotation,making weakly supervised methods even more important.In addition,in the real world,3D models often contain different rotational poses,and rotational invariance of the processing results becomes an important challenge.To address the problems of weakly supervised part segmentation and rotation resistant classification of 3D model using a sphere node graph representation of 3D models,this paper presents a targeted analysis and study described as follows:(1)This paper investigates the role of the sphere node graph representation of 3D models in3 D model segmentation tasks and proposes a weakly supervised part segmentation method for3 D models based on sphere node graph.For a given 3D model(usually in the form of a mesh),this paper extracts the sphere node graph from its interior as a discriminator for point cloud segmentation.By training the sphere nodes with segmentation properties consistent with the small number of labelled points,the segmentation results of the surface points close to the sphere nodes in the test model can be predicted.The individual segmentation labels of the training sphere nodes can initially achieve weakly supervised learning,but each model contains two to five different components,and the sphere nodes at the part junctions may belong to more than one different component.To overcome this problem,this paper achieves improved processing of the part connections by predicting the tangent plane normal vectors parametrized by the center of the sphere and the bipartition labels.It is shown that this significantly improves the accuracy of the surface point cloud segmentation at the part joints,and that our approach achieves an overall accuracy similar to that of strongly supervised methods(using more than 1K labelled points).(2)This paper investigates the anti-rotation feature extraction of the sphere node graph representation of 3D model and applies it to the anti-rotation classification and part segmentation tasks of 3D model.The positions of the sphere nodes in the sphere node graph are rotated with the center of the model as a whole,so the distance,angle and connected edge information between them are rotation invariant.Based on this,a rotation invariant feature representation for the sphere nodes is designed in this paper.This form of feature representation requires only one training sample under any unknown rotation for each training model,and the same feature representation can still be obtained after any rotation for the test model,thus requiring a weakly supervised form with few samples.Experimental results demonstrate that this method has better anti-rotation classification and part segmentation performance than traditional methods,and can efficiently enhance the accuracy and stability of 3D model classification and segmentation tasks.