An official implementation for Splatfacto-W.

Please update your nerfstudio to the latest development version.

Ensure that nerfstudio has been installed according to the instructions. Clone or fork this repository and run the commands:

conda activate nerfstudio
cd splatfacto-w/
pip install -e .
ns-install-cli

You can download the phototourism dataset from running.

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

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:

|---brandenburg_gate
|   |---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 [OPTIONS] --data [PATH] [dataparser]

You can add these options to enhance the training process:

1. --pipeline.model.enable_bg_model True

   • Enables background modeling
   • Useful for scenes with distinct foreground and background

2. --pipeline.model.enable_alpha_loss True

   • Enables alpha loss for punishing gaussians from occupying sky areas.
   • Should be used with background modeling

3. --pipeline.model.enable_robust_mask True

   • Improves handling of transient objects (things that appear/disappear between views)

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.