8000 GitHub - Sunnyhong0326/DA-GS: A re-implementation of decoupled appearance for 3D Gaussian Splatting
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Sunnyhong0326/DA-GS

 
 

Repository files navigation

Decoupled Appearance for 3D Gaussian Splatting

This repository aims to reproduce the decoupled appearance modules from multiple papers including

Motivation

To train a high quality radiance field without floaters, we often need to carefully capture the scene by fixing the exposure, white balance and other camera parameters to ensure the assumption of multiview consistency. However, for in-the-wild captures, it is hard to maintain the multiview consistency assumption because of varying lightning. Therefore, the radiance field would appear many floaters.

Many papers handles this kind of varying lightning condition captures by introducing neural network into training 3DGS. Papers include:

Although these methods are good at modeling appearance variations by using neural network, we need to inference a neural network at test time which may cause FPS drop when the scene is larger.

We seek for different approaches by decoupling the appearance variations and fuse into a floater-free radiance field which can be used in a normal rendering pipeline without inferencing a neural network.

TODOs

Support multiple decoupled appearance modules

Installation

Docker

You can build docker image from scratch or pull the docker image

Build from scratch

docker build -t dags .

Pull docker image

docker pull sunnyhong/dags

Run container

docker run -it --gpus all -v your_dags_path:/app -v your_dataset_path:/app/data -v your_output_path:/app/output bash
# Run code in the container
conda activate da-gs
python train.py -s data/dataset_name -m output/dataset_name

Please refer to the original repository gaussian-splatting for real-time viewer usage

About

A re-implementation of decoupled appearance for 3D Gaussian Splatting

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 99.3%
  • Dockerfile 0.7%
0