VastGaussian

What is VastGaussian?

VastGaussian is an open-source project for 3D scene reconstruction that uses 3D Gaussians to model large-scale scenes. This project was developed from scratch by the author, and while it may contain some errors, it introduces a novel approach to 3D scene reconstruction. Its key strengths include its ability to handle large datasets and improvements over the original 3DGS project, making it more accessible and easier to use.

Target Audience:

This project is ideal for researchers and developers working in the fields of 3D scene reconstruction, computer vision, and graphics. VastGaussian offers a new method for 3D scene reconstruction, capable of processing large datasets and providing improvements over the original 3DGS project, making it more understandable and usable.

Example Scenarios:

Application on the UrbanScene3D dataset

Application on the Mill-19 dataset

Application on the tandt_db dataset

Key Features:

Camera-position-based region division

Position-based data selection

Visibility-based camera selection

Coverage-based point selection

Decoupled Appearance Modeling

Seamless Merging

Parallel training of multiple regions on a single GPU

Getting Started:

1. Clone or download the VastGaussian project to your local machine.

2. Set up the environment as per the project documentation, including installing necessary dependencies.

3. Adjust parameters in the arguments/parameters.py file to fit your dataset and requirements.

4. Start training the VastGaussian model using the train_vast.py file.

5. Use provided scripts or command-line tools to evaluate and utilize the model.