Reproduce “BendTwin: Robust Dense-to-Sparse Physical Reconstruction with Bending-Aware Differentiable Spring-Mass Models” on OpenResearch

About this paper

Yixiong Jing, Qi Wang, Lin Chen, Junwei Jiang, Guangming Wang, Haibing Wu, Olaf Wysocki, Wanli Ma and 1 more

BendTwin introduces bending stiffness and damping into differentiable spring-mass models for robust physical reconstruction, creating more mechanically faithful digital twins from sparse-view RGB-D videos. This framework achieves a 19.0% reduction in Chamfer Distance during reconstruction and a 16.8% reduction for future prediction, particularly enhancing stability with sparser physical representations.

Key insights

• Explicitly modeling bending mechanics significantly improves the mechanical fidelity and stability of spring-mass systems, especially when the physical graph discretization is sparse.
• Bending constraints effectively penalize non-physical surface distortions, enabling purely surface-based models to reproduce volumetric deformations with comparable or superior accuracy to axial-only systems with interior nodes.
• The differentiable formulation of bending forces allows for robust inverse optimization of mechanical parameters from observational video data, enhancing the accuracy of reconstructed digital twins.

Summary by alphaXiv.