OSMGE: One-Shot Multiscale Geometric Encoding for Texture Segmentation in 3D Meshes

Mapping textures to surfaces with precise facet-level classification significantly enhances the realism of 3D scenes. Identifying texture and non-texture regions at this granular level improves the authenticity and quality of computer graphics and animations. This paper focuses on classifying facets on a 3D mesh into texture and non-texture categories, analogous to pixel-level segmentation in 2D images. Each facet is initially encoded into a 2D feature map, using its neighbors' localized geometric properties within a predefined radius. Given that 3D meshes vary in resolution, a multiscale version of the feature maps is employed to ensure the model's robustness. These feature maps are input into an ensemble framework of multiple simple convolutional neural network (CNN) models, enhancing predictive capabilities. A key highlight of the proposed approach is its ability to effectively generalize from a single 3D sample, making it suitable for scenarios with limited data. The predictions of each model in the ensemble are weighted and combined for the final inference. The proposed model is rigorously evaluated on the SHREC'18, SHREC'17, and KU 3DTexture datasets, demonstrating state-of-the-art performance. (Code will be released upon acceptance.)