Reproduce “MeshLoom: Feed-Forward Non-Rigid Registration of Mesh Sequences” on OpenResearch

About this paper

Jianqi Chen, Jiraphon Yenphraphai, Xiangjun Tang, Sergey Tulyakov, Chaoyang Wang, Peter Wonka, Rameen Abdal

Researchers introduced MeshLoom, a feed-forward neural network for non-rigid registration of variable-length mesh sequences, producing topology-consistent, explicit per-vertex deformations. The model achieves 50% faster inference and over 10% higher DreamSim scores compared to prior methods while generalizing across diverse object categories and enabling motion interpolation.

Key insights

• Encoding topology-aware point representations for the anchor mesh is critical to prevent vertex entanglement and ensure coherent deformation, even for geometrically close but topologically distant regions.
• Decomposing total vertex displacement into a global centroid shift and local residual deformation enhances robustness to large translations and yields perceptibly smoother mesh surfaces.
• A multi-modal transformer architecture, integrating 3D shape priors and optional image features, effectively learns inter-frame correspondences and global motion embeddings across diverse object categories.

Summary by alphaXiv.