Stylization of Photographs using Tractography

Using the concepts from diffusion tensor imaging (DTI) and image processing, this project aims to stylize photographs in painterly style.

For details, read the preprint here.

Usage

Install uv and run:

uv run main.py INPUT_IMAGE OUTPUT_IMAGE [--params PARAMS_JSON] [--orientation-vector {structural,gradient}]

The optional PARAMS_JSON file can be used to specify parameters for the algorithm. It should be a JSON file with the following structure:

[
    { // First layer
         "sigma": 1.0,               // Standard deviation for Gaussian smoothing
         "length_lines": 10000.0,    // Maximum length of lines in this layer
         "width": 2.0,               // Width of lines in this layer
         "min_length": 10.0,         // Minimum length of lines to draw
         "color_threshold": 50.0     // Minimum color difference to draw a line
    },
    { // Second layer
        ...
    },
    ...
]

The orientation-vector argument specifies how to compute the orientation. If structural it uses the primary eigenvector of the structure tensor (proposed improvement), if gradient it uses the image gradient (for reference).

The settings used in the preprint can be found in test-images/params.json.

Reproducibility

To reproduce the results in the preprint, run the following commands:

# Fig 1
uv run main.py test-images/cubain.jpg out_cubain.png --params test-images/params.json --orientation-vector structural

# Fig 3
uv run main.py test-images/picasso.jpg picasso_st.png --params test-images/params_st_vs_gt.json --orientation-vector structural
uv run main.py test-images/picasso.jpg picasso_gt.png --params test-images/params_st_vs_gt.json --orientation-vector gradient

# Stylization of all images in the test datasets (Figures 4-11)
source .venv/bin/activate  # MacOS/Linux
& .\.venv\Scripts\Activate.ps1  # Windows PowerShell
cd test-images
bash run_all.sh # MacOS/Linux
& run_all.ps1 # Windows PowerShell
bash make_montage.sh # MacOS/Linux (requires ImageMagick)