Reproduce “Tessellation GS: Neural Mesh Gaussians for Robust Monocular Reconstruction of Dynamic Objects” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shuohan Tao, Boyao Zhou, Hanzhang Tu, Yuwang Wang, Yebin Liu
Tessellation GS introduces a self-supervised framework for robust monocular 3D reconstruction of general dynamic objects. It employs a novel structured 2D Gaussian representation anchored to a deforming mesh, achieving state-of-the-art photorealistic rendering and precise geometry while mitigating view overfitting on dynamic benchmarks.
Key insights
- Integrating coarse geometric priors from LRMs significantly enhances the robustness of monocular dynamic reconstruction by providing initial grounding and reducing ambiguity.
- Structuring 2D Gaussians on a deforming mesh, combined with explicit constraints on their scale and offset, effectively mitigates view overfitting and promotes geometrically consistent reconstructions.
- The hierarchical Mesh-Gaussian Quad Tree, with its learnable subdivision and competitive opacity mechanism, enables adaptive allocation of Gaussian density, allowing for fine-detail modeling while maintaining efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.