Research On 3D Model Retrieval And Classification In Open Environment

In recent years,the 3D model has been widely used in many fields,and the amount of data has increased dramatically,resulting in a large number of 3D model data sets.However,only a few data sets for academic research have complete label information.In an open environment,due to the high cost of manual labeling,most of the 3D model data sets usually lack label information,which is generally expressed in two scenarios:1)Unlabeled scenario,that is,there are no labeled samples in the data set;2)Few-shot scenario,that is,there are extremely few labeled samples in each category.Traditional 3D model retrieval and classification methods rely on a large number of labeled samples in a specific closed data set,which results in a lack of generalization ability and cannot be directly applied to unlabeled 3D models.How to realize the retrieval and classification of 3D models in unlabeled or few-shot scenarios has become an urgent problem to be solved.To this end,this article focuses on the problems of unlabeled 3D model retrieval and few-shot 3D model classification in an open environment,and has carried out the following two aspects:(1)Aiming at the problem of unlabeled 3D model retrieval,a cross-domain 3D model retrieval method based on domain separation and pseudo-label is proposed.This method combines 3D model visual feature learning and unsupervised domain adaptive learning and can align the source domain of the richly labeled 3D model with the target domain of the unlabeled 3D model,and realize the unlabeled retrieval of the target domain 3D model.This method uses a domain-specific batch normalization module to separate the batch statistical characteristics of the source domain and the target domain to reduce the interference of domain-specific information;at the same time,based on semantic centroid alignment,the target domain pseudo-labels are generated through two-stage iterative training to achieve Feature alignment at the category level between domains.Through comparative experiments and ablation experiments on the public data sets NTU and PSB,the superiority of this method is verified.(2)Aiming at the problem of few-shot 3D model classification,a 3D model classification method based on meta-learned multi-modal fusion is proposed.This method combines the multi-modal representation learning of the 3D model with metalearning.By training with limited labeled samples,a better classification performance can be achieved.This method is based on meta-learning paradigm training,and uses the multi-modal representation fusion module to fuse the point cloud and multi-view information of the 3D model to generate a richer representation;at the same time,it uses neuron-level scaling and shifting operations to transfer prior knowledge and reduce parameter scale of the network.Through comparative experiments and ablation experiments on two derived data sets,Meta-ModelNet and Meta-ShapeNet,the superiority of this method is verified.