A Recipe for Geometry-Aware 3D Mesh Transformers
Mohammad Farazi, Yalin Wang
Summary
Utilizing patch-based transformers for polygon meshes poses challenges due to the lack of canonical ordering and input size variations. Our study introduces a geometry-aware 3D mesh transformer with novel components like spectral-preserving tokenization, geodesic masking, and heat diffusion-based embeddings, demonstrating superior performance and efficiency in 3D segmentation and classification tasks.
Figures (click on each for a larger version):
