V2.0-alpha

Changelog

This update includes a series of major changes to speed up the forward and backward passes, improve the optimization quality, simply the codebase structures. In particular, the code is fundamentally restructured to work with pytorch3d.

Key features

• Forward rendering
  • speedup with uniform grid neighborhood search FRNN.
  • improved rendering quality based using anisotropic 3D gaussian Vrk_isotropic: False
• Inverse rendering
  • more stable and faster convergence by disentangling the inverse rasterization from color information using silhouette_loss.
  • speedup from coalesced memory access on GPU.
  • projection and repulsion terms as loss terms for convenient control.
  • allow Adam optimization and joint normal and point position optimization.

Todos

[ ] add back example for point surface editting (e.g. denoising)
[ ] update to latest pytorch and pytorch3d versions