3D Model Classification Based on Contrastive Learning

At present, 3D model classification has been a research hotspot. Massive 3D models not only have diversity in each class, but also have similarities between classes, which seriously affect the classification accuracy of 3D models. We propose a 3D model classification method based on contrastive learning. In this method, the training is divided into a sample discrimination stage and a classification stage. In the stage of sample discrimination, 3D models of the same category are mutually positive samples, and 3D models of other categories are mutually negative samples. The contrastive loss is used to constrain the sample features, and the positive and negative samples are mapped to the single-center unit hypersphere in the same space to obtain a good semantic representation space of 3D model classification. In addition, in order to capture the correlation between the views and the key areas in the views, a multi head self-attention module and spatial attention module are introduced in the paper. Moreover, the channel attention is added in the multi-head self-attention module to obtain the channel dimension information. In the classification stage, the network model is transferred to the classification task by fine-tuning the network parameters to complete the 3D model classification. The experimental results show that the classification accuracy of the 3D model respectively reaches 99. 4% and 97. 5% on the ModelNet10 and ModelNet40 datasets.