Gaussian Splatting and Mesh Reconstruction
This is the digital version of my thesis, Geometric Mesh Reconstruction and Differentiable Smoothing for Gaussian Splatting. It walks through how a scene captured with photos becomes a clean, low-polygon 3D mesh, and describes my two contributions:
- Meshfacto — porting the state-of-the-art Mesh-In-the-Loop (MILo) method into the Nerfstudio library, cutting the polygon count of the extracted meshes dramatically.
- Differentiable smoothers — extra training losses that regularize the mesh and stabilize the optimization.
