In this project I will be using 3 inpainting methods on MIT Places Dataset. 2 inpainting methods are based on neural networks. On the other hand, 1 method based on statistics.
For this approach I worked with the offical code and design. So I won't add the same codes to this repository. You can check code files from their Github and follow the steps for training. My test results and test codes will be under this directory.
For this research, Nvidia did not share all the code files for inpainting. So I worked with unofficial repository, MathiasGruber's PConv-Keras repo.. I made changes on Step 4 (Training part) and I rewrote Test part. If you work with 256, 256 images like me, you should change the img_rows and img_cols variables in the libs/pconv_model.py line 21. The code files that I changed and test results will be under this directory.
This method is easy to work, because there is no training process. You have to use Python2. In Windows I encountered so many erors about PyMaxflow in the installation process, it was related to Visual C++. So, I used Ubuntu, it was much smooth. I edited code little bit, to work with all images under a directory instead of a single image. Code file is under this directory. You must make changes in config.py. Code is very interwined with this file. So I didn't break this connection between them.
If you have any question about implementation of this methods don't hesitate to ask me. Good luck! 🌠