Advancing 3D Mesh Segmentation with Regularization Losses: A Semi-Supervised Approach Using MeshCNN

The application of deep learning in 3D mesh segmentation has attracted increasing attention. In supervisedlearning based mesh segmentation, back propagation is often based on cross entropy loss. Some prior knowledge, such as the structure similarity or geometric consistency of data, can be used to improve the segmentation effect. However, in the literature for mesh segmentation, such prior knowledge is usually either ignored in deep learning or only considered at post-processing stage. In this work, we add prior knowledge as a regularization term into the loss function for optimization, which is applied in the semi-supervised 3D mesh segmentation based on MeshCNN for the first time. A large number of experiments show that, for both semi-supervised segmentation with 30% and 70% labeled data and fully supervised segmentation, the proposed regularization term improves the segmentation performance.