3D-Gaussian-Compression-Leveraging-Clustering-Algorithm

As a representation method for displaying 3D scenes, 3D Gaussian allows optimization with state-of-the-art (SOTA) visual quality and competitive training times. In comparison to previous explicit scene representations, although the anisotropic Gaussians are capable of modelling complex shapes with a lower number of parameters, it still requires a large amount of memory and storage.

This project is a faithful PyTorch implementation of 3DGS that reproduces the results while the storage volume is reduced to 0.24 times the original size. The code is based on authors' implementation here, and has been tested to match it numerically.

Overview

The codebase has 3 main components:

Original code implementation
Clustering 3DGS points using K-Means
Clustering 3DGS points using MeanShift

How to Run?

train a vanilla Gaussian

python train.py -s <path to COLMAP or NeRF Synthetic dataset> -m <path to store the trained model>--eval

After training for 30k iterations (~20 minutes on a single 4090), you can find the following pointcloud at

/data/output/point_cloud/iteration_30000/point_cloud.ply

This position depends on the path to store the trained model. The path used here is

python train.py -s data -m data/output --eval

Clustering 3DGS points using K-Means

After getting the vanilla Gaussian points, you need to run kmeans.py

python kmeans.py --file <path to store Gaussian point cloud> --eval <whether to test>

Then, in the same path, you can find the following pointcloud at

/data/output/point_cloud/iteration_30000/point_cloud.ply

Finally, if you need to render the image, you need to run render_vq.py

python render.py -m data/output --eval

Clustering 3DGS points using MeanShift