Better Gaussian Splatting in Julia

GaussianSplatting.jl 2.0 has been released with significant improvements in performance and usability.

Key Updates:

• Multi-GPU Backend Support: Supports various GPU environments such as NVIDIA (CUDA), AMD, and MacBook (Metal) via KernelAbstractions.jl, allowing execution on multiple targets with a single codebase.
• Multithreaded UI: Separates the UI (Frontend) from rendering/training (Backend), ensuring the UI remains responsive during data loading or kernel compilation, with real-time Loss plots and progress updates.
• MCMC Densification Strategy: Introduces the MCMC (Markov Chain Monte Carlo) strategy alongside existing methods to reduce initialization dependency and precisely control the number of Gaussians.
• Depth and Geometric Supervision: Improves reconstructed geometry and reduces floaters using depth information from tools like Depth-Anything 3. Additionally, normal consistency and surface flattening features allow for precise control over surface shapes.
• Sky Dome Feature: Supports a Sky Dome to separate background from the scene, resolving issues where sky/background elements affect scene geometry and reducing floaters.

Other additions include automatic checkpoint saving, hyperparameter configuration file support, and camera path reproduction.