SAMesh is a novel zero-shot method for mesh part segmentation that addresses the limitations of traditional shape analysis (e.g. Shape Diameter Function (ShapeDiam)) and learning-based approaches. It operates in two phases: multimodal rendering, where multiview renders of a mesh are processed through Segment Anything 2 (SAM2) to generate 2D masks, and 2D-to-3D lifting, where these masks are combined to produce a detailed 3D segmentation. Compared to other 2D-to-3D lifting methods, SAMesh does not require an input vocabulary, which limits those methods to semantic segmentation as opposed to part segmentation. SAMesh demonstrates good performance on traditional benchmarks and superior generalization on a newly curated dataset of diverse meshes, which we release below.
Examples of running SAMesh on our curated dataset:
Pipeline of SAMesh:
To install SAMesh, use the following commands:
pip install -e .Don't forget to init the submodules and pip install -e on them respectively. We tested SAMesh on python 3.12 and cuda 11.8. If you encounter issues with building SAM2, try with the --no-build-isolation flag. If you pyrenderer issues related to ctypes, try installing PyOpenGL==3.1.7.
Download a SAM2 checkpoint as provided in the SAM2 repo. notebooks/mesh_samesh.ipynb and notebooks/mesh_shape_diameter_function.ipynb are detail how to setup and run SAMesh and ShapeDiam, respectively. Some mesh examples from the curated dataset are provided in assets.
George Tang*, William Zhao, Logan Ford, David Benhaim, Paul Zhang
*Work done during an internship at Backflip AI.