Splatfacto in the Wild#

This is the implementation of Splatfacto in the Wild: A Nerfstudio Implementation of Gaussian Splatting for Unconstrained Photo Collections. The official code can be found here.


This repository follows the nerfstudio method template

1. Install Nerfstudio dependencies#

Please follow the Nerfstudio installation guide to create an environment and install dependencies.

2. Install the repository#

Run the following commands: pip install git+https://github.com/KevinXu02/splatfacto-w

Then, run ns-install-cli.

3. Check installation#

Run ns-train splatfacto-w --help. You should see the help message for the splatfacto-w method.

Downloading data#

You can download the phototourism dataset from running.

ns-download-data phototourism --capture-name <capture_name>

Running Splafacto-w#

To train with it, download the train/test tsv file from the bottom of nerf-w and put it under the data folder (or copy them from ./splatfacto-w/dataset_split). For instance, for Brandenburg Gate the path would be your-data-folder/brandenburg_gate/brandenburg.tsv. You should have the following structure in your data folder:

|   |---dense
|   |   |---images
|   |   |---sparse
|   |   |---stereo
|   |---brandenburg.tsv

Then, run the command:

ns-train splatfacto-w --data [PATH]

If you want to train datasets without nerf-w’s train/test split or your own datasets, we provided a light-weight version of the method for general cases. To train with it, you can run the following command

ns-train splatfacto-w-light --data [PATH] [dataparser]

For phototourism, the dataparser should be colmap and you need to change the colmap path through the CLI because phototourism dataparser does not load 3D points.