Sort-free Gaussian Splatting via Weighted Sum Rendering

This repository is a re-implementation of the ICLR'25 paper "Sort-free Gaussian Splatting via Weighted Sum Rendering". Paper.

We are honored that this repository has been recognized and linked as the released reference code by the author, Qiqi Hou.

Our implementation builds upon the original 3DGS by incorporating global parameters sigma and weight_background, as well as a set of sh parameters for each GS to decode the view-dependent effects Vi.

Performance on MipNeRF360

PSNR

	bicycle	bonsai	counter	flowers	garden	kitchen	room	stump	treehill	AVG
Paper	24.20	32.05	29.53	20.45	27.78	31.38	31.93	25.39	22.01	27.19
Ours	23.89	33.20	29.96	20.62	27.07	31.51	32.35	24.79	21.49	27.21
Ours-w/o D	23.89	33.05	29.83	20.78	27.14	31.21	31.65	24.90	21.61	27.12

SSIM

	bicycle	bonsai	counter	flowers	garden	kitchen	room	stump	treehill	AVG
Paper	0.744	0.938	0.909	0.580	0.872	0.923	0.925	0.728	0.614	0.804
Ours	0.714	0.951	0.919	0.581	0.850	0.931	0.930	0.708	0.620	0.800
Ours-w/o D	0.714	0.948	0.918	0.582	0.851	0.931	0.929	0.713	0.622	0.801

LPIPS

	bicycle	bonsai	counter	flowers	garden	kitchen	room	stump	treehill	AVG
Paper	0.205	0.199	0.191	0.342	0.097	0.125	0.197	0.235	0.311	0.211
Ours	0.265	0.144	0.176	0.329	0.128	0.117	0.190	0.275	0.347	0.219
Ours-w/o D	0.265	0.175	0.176	0.329	0.128	0.117	0.190	0.272	0.344	0.222

Ours - w/o D means not adding the --depth_correct parameter during training.

We have added a new parameter --depth_correct to train.py, which controls the effect of Equation 10 in the paper. If this parameter is included, the weight calculation is performed as follows:

$$ w(d_i) = \max\left(0, 1 - \frac{d_i}{\sigma}\right) v_i $$

which is consistent with the original paper. If the parameter is not included, the weight calculation change to:

$$ w(d_i) = v_i $$

This modification allows us to experiment with the impact of depth correction on the rendering process and compare the results with the original model.

Install

We tested on a server configured with Ubuntu 22.04, CUDA 11.8.

1. Clone this repo:

git clone --recursive https://github.com/LiYukeee/sort-free-gs.git --recursive
cd sort-free-gs

2. Install dependencies

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

Data

First, create a data/ folder inside the project path by

mkdir data

The data structure will be organised as follows:

data/
├── dataset_name
│   ├── scene1/
│   │   ├── images
│   │   │   ├── IMG_0.jpg
│   │   │   ├── IMG_1.jpg
│   │   │   ├── ...
│   │   ├── sparse/
│   │       └──0/
│   ├── scene2/
│   │   ├── images
│   │   │   ├── IMG_0.jpg
│   │   │   ├── IMG_1.jpg
│   │   │   ├── ...
│   │   ├── sparse/
│   │       └──0/
...

The MipNeRF360 scenes are provided by the paper author here.

Training

In our repository, the parameter settings for train.py, render.py, and metrics.py are identical to those of 3DGS.

MipNeRF360

bash train_mipnerf360.sh

Single Scene

python train.py --eval -s <path to COLMAP or NeRF Synthetic dataset> -m <path to output> --depth_correct
python render.py -m <path to output>
python metrics.py -m <path to output>

LICENSE

Please follow the LICENSE of 3D-GS.