For learning purposes, we offer an unofficial implementation of 3D Gaussian Splatting for Real-Time Radiance Field Rendering. Additionally, we provide a set of tools and documentation to facilitate the study of 3D Gaussian Splatting.
We want our implementation to be easy to use, the code to be easy to read, and the documentation to be easy to understand. Therefore, we named our project "EasyGaussianSplatting".
- Detailed documentation to demonstrate the mathematical principles of 3D Gaussian Splatting
-
Based on our documentation, re-implement 3D Gaussian Splatting.
- Forward on CPU
- Forward on GPU
- Backward on CPU
- Backward on GPU
- Full training implemention.
-
Provide tools for learning 3D Gaussian Splatting.
- A efficient viewer based on pyqtgraph for showing 3D Gaussian data (trained model).
- A demo showing how spherical harmonics work.
- Verify all computation processes of backward using numerical differentiation.
pip3 install -r requirements.txt
pip3 install gsplatcu/.
Given camera information, render 3D Gaussian data onto the 2D image.
The TRAINED_FILE
is the .ply file generated by the official Gaussian Splatting, or the .npy file generated by our train.py.
CPU version
python3 forward_cpu.py --gs='THE_PATH_OF_YOUR_TRAINED_FILE'
GPU version.
python3 forward_gpu.py --gs='THE_PATH_OF_YOUR_TRAINED_FILE'
Download the T&T+DB COLMAP datasets. T&T+DB COLMAP (650MB)
python3 train.py --path='THE_PATH_OF_UNZIPPED_DATASET'
(The Adaptive Density Control has not been implemented yet.)
A efficient 3D Gaussian splatting viewer for showing 3D Gaussian data.
python3 gaussian_viewer.py --gs='THE_PATH_OF_YOUR_TRAINED_PLY_OR_NPY_FILE'
A demo showing how spherical harmonics work.
python3 sh_demo.py
"The ground truth Earth image is modified from URL. By Solar System Scope. Licensed under CC-BY-4.0"
Verify all computation processes of backward using numerical differentiation.
python3 backward_cpu.py
python3 backward_gpu.py