Frequency-based View Selection in Gaussian Splatting Reconstruction

Monica M.Q. Li, Pierre-Yves Lajoie, and Giovanni Beltrame

This repository is code for the associated with the paper "Frequency-based View Selection in Gaussian Splatting Reconstruction", which can be found here. Please install according to the following steps. We only tested the code on Ubuntu.

Installation

Clone this repo with the flag --recursive.

Follow the steps in 3D-GS to create a conda environment and install the packages in environment.yml, as well as the submodules diff-gaussian-rasterization and simple-knn.

Please note that if you are using Ubuntu 24, you need to downgrade gcc and g++ to 9 to compile the submodules.

Install colmap and glomap.

Download the datasets and extract them in ./tandt_db/. The structure of the dataset directory should be like:

|---tandt_db
|   |---tandt
|       |---train
|           |---sparse
|               |---0
|                   |---images.bin
|                   |---points3D.ply
|                   |---points3D.bin
|                   |---cameras.bin
|                   |---project.ini
|           |---images
|               |---000081.jpg
|               |---000103.jpg
|               |---000240.jpg
|               |---000157.jpg
 ...
|       |---truck
...
|   |---db
|       |---drjohnson
...
|       |---playroom
...

Run

python3 gs-fft.py

Parameters

Lines 322-330 contain the parameters for input and output paths, iterations to train the Gaussian models and the iterations to select the next-best-views.

Lines 95-100 contain the parameters the same as the ones in "Command Line Arguments for train.py" in 3D-GS.

Citation

BibTeX

@article{li2024frequency,
  title={Frequency-based View Selection in Gaussian Splatting Reconstruction},
  author={Li, Monica MQ and Lajoie, Pierre-Yves and Beltrame, Giovanni},
  journal={arXiv preprint arXiv:2409.16470},
  year={2024}
}