English | 中文

Introduction

We provide a training method based on the original 3DGS and an effective post-processing solution, so that everyone can train in the simplest way. Of course, you can also modify the parameters to achieve the results you want.

Installation

conda env create --file environment.yml
conda activate gaussian_splatting

cd submodules/
python ./diff-gaussian-rasterization/setup.py install
python ./simple-knn/setup.py install
python ./fused-ssim/setup.py install

Training

We try to simplify the input parameters of training. You only need to provide a video or a set of pictures to complete the training directly through the following script.

cd gaussian-splatting
# fps=0.5 means that the video takes one frame every two seconds
python run.py --video_path /path/of/video --fps 0.5
or
python run.py --image_path /path/of/images

Training speed acceleration

Since our method is based on the original gaussian-splatting, we also support training acceleration.

--optimizer_type sparse_adam

Postprocess

To reduce outliers and floaters to improve rendering quality, We post-processed it through 3dconverter. It is worth noting that this method is more effective in removing floaters of objects, but it is easy to accidentally delete the main structure of the scene. Therefore we recommend that you use this post-processing method only for objects.

--post_process

Program Packaging

We provide a packaging program in pack.py, you can package the entire process from video or image to 3DGS model into an executable file.

python pack.py