Mini-Splatting2: Building 360 Scenes within Minutes via Aggressive Gaussian Densification

This is the official implementation of Mini-Splatting2, a point cloud reconstruction work in the context of Gaussian Splatting. Through aggressive Gaussian densification, our algorithm enables fast scene optimization within minutes. For technical details, please refer to:

Mini-Splatting2: Building 360 Scenes within Minutes via Aggressive Gaussian Densification
Guangchi Fang and Bing Wang.
[Paper]

(1) Setup

This code has been tested with Python 3.8, torch 1.12.1, CUDA 11.6.

• Clone the repository

git clone [email protected]:fatPeter/mini-splatting2.git && cd mini-splatting2

• Setup python environment

conda create -n mini_splatting2 python=3.8
conda activate mini_splatting2
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

• Download datasets: Mip-NeRF 360, T&T+DB COLMAP.

(2) Mini-Splatting2 (Sparse Gaussians)

Training scripts for Mini-Splatting2 are in msv2:

cd msv2

• Train with train/test split:

# mipnerf360 outdoor
python train.py -s <dataset path> -m <model path> -i images_4 --eval --imp_metric outdoor --config_path ../config/fast
# mipnerf360 indoor
python train.py -s <dataset path> -m <model path> -i images_2 --eval --imp_metric indoor --config_path ../config/fast
# t&t
python train.py -s <dataset path> -m <model path> --eval --imp_metric outdoor --config_path ../config/fast
# db
python train.py -s <dataset path> -m <model path> --eval --imp_metric indoor --config_path ../config/fast

• Modified full_eval script:

python full_eval.py -m360 <mipnerf360 folder> -tat <tanks and temples folder> -db <deep blending folder>

(3) Mini-Splatting2-D (Dense Gaussians)

Training scripts for Mini-Splatting2-D are in msv2_d:

cd msv2_d

• Train with train/test split:

# mipnerf360 outdoor
python train.py -s <dataset path> -m <model path> -i images_4 --eval --imp_metric outdoor --config_path ../config/fast
# mipnerf360 indoor
python train.py -s <dataset path> -m <model path> -i images_2 --eval --imp_metric indoor --config_path ../config/fast
# t&t
python train.py -s <dataset path> -m <model path> --eval --imp_metric outdoor --config_path ../config/fast
# db
python train.py -s <dataset path> -m <model path> --eval --imp_metric indoor --config_path ../config/fast

• Modified full_eval script:

python full_eval.py -m360 <mipnerf360 folder> -tat <tanks and temples folder> -db <deep blending folder>

(4) Dense Point Cloud Reconstruction

This implementation directly support dense point cloud reconstruction:

# similar to train.py (-i images_4/images_2, --imp_metric outdoor/indoor)
# output ply files are saved in ./teaser
python teaser.py -s <dataset path> -m <model path> -i images_4 --eval --imp_metric outdoor

Acknowledgement. This project is built upon Mini-Splatting, 3DGS and Taming 3DGS.