TartanAir to NerfStudio Converter

TartanToNerfStudio is a Python script that converts the TartanAir and TartanGround dataset format into the NerfStudio dataset format. It processes pose files, converts depth images (if needed), and generates a transforms.json file compatible with NerfStudio.

Features

• Reads pose data from TartanAir datasets.
• Supports three camera intrinsic configurations: Air, Ground, and Custom.
• Converts pose data from TartanAir format to NerfStudio's 4x4 transformation matrices.
• Optionally limits the number of poses and distributes them uniformly.
• Converts depth images to .npy format if required.
• Outputs a transforms.json file compatible with NerfStudio.

Installation

1. Clone this repository:

   git clone https://github.com/IQisMySenpai/tartanToNerfstudio.git
   cd tartanToNerfstudio

2. (Optional) Create a conda environment (Python 3.8): I recommend using your existing nerfstudio environment.

   conda create -n nerfstudio python=3.8

3. Activate the environment:

   conda activate nerfstudio

4. Install the required dependencies:

   pip install -e .

Usage

Command-Line Interface

Run the script using tyro CLI:

tartanToNerfstudio <base_path> [--pose-limit N] [--uniform] [--spatial-uniform] [--cube-size SIZE] [--depth-conversion] [--camera-intrinsics {Air,Ground,Custom}] [--output-path OUTPUT]

Arguments:

Argument	Type	Required	Description
base_path	Path	Yes	Path to the base folder of the TartanAir dataset.
--pose-limit, -p	int	No	Limit the number of poses to convert.
--uniform, -u	bool	No	Distribute selected poses uniformly along the trajectory instead of taking the first n poses (Default: False).
--spatial-uniform, -su	bool	No	Distribute poses spatially uniformly in 3D space using a cube subdivision of the scene (Default: False).
--cube-size, -cs	float	No	Edge length (in meters) of the cubes used for spatial subdivision when --spatial-uniform is enabled (Default: 10.0).
--depth-conversion, -d	bool	No	Convert depth images to .npy format (Default: False). Needed for Tartan Ground.
--camera-intrinsics, -c	Air/Ground/Custom	No	Camera intrinsics to use: Air, Ground, or Custom (Default: Air).
--output-path, -o	Path	No	Output transforms file path (default: transforms.json in base folder).

Spatially Uniform Pose Sampling

When --spatial-uniform is enabled (and --pose-limit is set), all poses are first collected and then sampled so that the selected cameras are spread out uniformly in 3D space:

• The scene bounds are estimated from all camera positions.
• The space is subdivided into cubes of size --cube-size meters.
• Each cube that contains cameras contributes a roughly equal number of poses (up to the requested limit), ensuring good spatial coverage of the scene.

Note that --uniform (temporal / index-based sampling) and --spatial-uniform are mutually exclusive; only one of them can be used at a time.

Example

Convert a TartanAir dataset with images and depth maps:

tartanToNerfstudio ./dataset --pose-limit 300 --camera-intrinsics Ground --depth-conversion

Note

In case you have a dataset with different camera intrinsics as the default tartanair datasets. You can change the camera_intrisincs variable in the tartanair_to_nerfstudio.py file to match your dataset.

Output

The script will generate a transforms.json file in the base folder with the following structure:

{
    "camera_model": "OPENCV",
    "fl_x": 320.0,
    "fl_y": 320.0,
    "cx": 320.0,
    "cy": 240.0,
    "w": 640,
    "h": 480,
    "frames": [
        {
            "file_path": "image0001.png",
            "transform_matrix": [[...]]
        }
    ]
}

License

This project is licensed under the MIT License. See LICENSE for details.

Contributing

Feel free to open issues or submit pull requests to improve this project. Contributions are welcome!