wip

docs/reading/papers/202311-sugar.md

@@ -0,0 +1,49 @@
+---
+keywords: gaussian splatting
+---
+
+# SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering
+
+precise and extremely fast mesh extraction from 3D Gaussian Splatting
+
+## Contributions
+
+- a regularization term that encourages the gaussians to align well with the surface of the scene
+- a method that exploits this alignment to extract a mesh from the Gaussians using Poisson reconstruction, which is fast, scalable, and preserves details, in contrast to the Marching Cubes algorithm usually applied to extract meshes from Neural SDFs
+- an optional refinement strategy that binds gaussians to the surface of the mesh, and jointly optimizes these Gaussians and the mesh through Gaussian splatting rendering
+
+---
+
+the gaussians are unstructured
+
+Gaussian Splatting performs densification in order to capture details of the scene with highfidelity... This results in a density function that is close to zero almost everywhere(?), and the Marching Cubes algorithm(?) fails to extract proper level sets(?) of such a sparse density function even with a fine voxel grid
+
+Instead, Poisson reconstruction algorithm, scalable
+
+## Related works
+
+light fields(?) Their work emphasized the importance of efficiently traversing volumetric data to produce realistic images(?)
+Traditional mesh-based IBR methods
+Volumetric IBR methods - NeRF
+Hybrid IBR methods
+Point-based IBR methods
+
+## Methods
+
+regularization term: derive an SDF from the gaussians, assuming (needs to reword)
+- well spread (the density of any position can be well approximated by only considering the closest gaussian)
+- flat (one of scaling factors close to 0)
+- opaque (alpha = 1)
+
+computing a slightly different loss relying on an SDF rather than on density further increases the alignment of gaussians with the surface of the scene
+the zero-crossings(?) of the Signed Distance Function
+
+We also add a regularization term to encourage the normals of SDF ¯f and the normals of SDF ˆf to also be similar(?)
+
+mesh extraction: ...
+
+refinement: instantiate a set of new thin 3D gaussians from each mesh triangle. The gaussians have only 2 learnable scaling factors and only 1 learnable 2D rotation
+
+## Results
+
+slightly(?) worse than vanilla 3DGS and Mip-NeRF360