Vanilla, Abs-GS, and MCMC densification
Bilateral grid (fully fused CUDA implementation), as well as another exposure correction method based on linear least squares
PhyGaussian and erank regularization to reduce spiky Gaussians
3DGUT for fisheye, with an MCMC-friendly way to eliminate large/spiky Gaussians for compatibility with vanilla 3DGS viewers
Use SH for background color, removes floaters from sky when combined with opacity regularization
Batching for very large scenes

Partially supports

Masking (sky mode and people/car mode)
Depth and normal supervision using monocular geometry models
Regularization to balance sky removal and discouraging transparency

TODO

Multi resolution loss
Better camera pose optimizer
Better mesh export
Faster training (Taming-GS backward, fused/lazy optimizer, Stop-the-pop tile culling, etc.)
Multi-GPU training

Additional features

Display number of Gaussians and training loss/PSNR/SSIM in terminal during training
Reduced system RAM usage for data loader when cached on CPU (up to 2x)

WebGL viewer features (see `webgl`)

Fisheye distortion (supports >180deg fisheye)
Compression
Collision detection (WIP)

Scripts (see `scripts`)

Extract frames from video, auto skip blurry frames
Segmentation using SAM-2
Process a folder of camera raw images to specific color space

Installation

Install Nerfstudio (see instructions). Clone this repository and run the commands:

cd spirulae-splat/
git checkout stable
git submodule update --init
MAX_JOBS=8 pip install -e . --no-build-isolation
ns-install-cli

Running the method

This repository creates a new Nerfstudio method named "spirulae". To train with it, run the command:

ns-train spirulae --data [DATASET_PATH]

By default, spirulae-splat uses all available images for training. To support ns-eval, for nerfstudio dataset, use the following command for training:

ns-train spirulae --data [DATASET_PATH] nerfstudio-data --train-split-fraction 0.9

README.md

spirulae-splat

Currently supports