Contour-guided Image Completion with Perceptual Grouping

Authors

Citation

If you use this code for your research, please cite our BMVC 2021 paper: Contour-guided Image Completion with Perceptual Grouping (citation coming soon).

Set-up

1. Install Python 3.8 on your computer
2. Install pip on your computer for managing python packages
3. Create and activate a virtual environment.
4. Install all package dependencies with pip install -r requirements.txt.

Project Structure

Running a simple example

To generate a Stochastic Completion Field (SCF) for two points as discussed in our BMVC 2021 paper, run the code in Stochastic_Completion_Fields_Pipeline.ipynb. In our paper, we integrate this SCF generating framework with image inpainting and denoising.

Random_Walks_Implementation

In this folder you will find a simple Monte-Carlo algorithm that generates approximate Stochastic Completion Fields using random walks. For details, see the included Random_Walks.pdf file inside this folder.

fokker_planck_experiment_runner

Our pipeline to generate Stochastic Completion Fields (SCFs) using the command line. This pipeline generates .npz files, which we integrated with EdgeConnect and Image Denoising algorithms to achieve the results in our paper.

Execution of pipeline

Inside the config folder, we provide a sample source and sink configuration for our SCF. Run the following command.

python3 experiment_runner.py --config_file configs/two_points.yml --experiment_dir experiments/two_points/

The results will be saved in the experiments directory that we specify. In this case, experiments/two_points/. In here, the completion field will be saved as a .npz file. You may use this .npz file in any application you wish.