This repository is part of the our proposed "Non-Lossy Ground Truth Comparison via Convolutional Auto-Encoders for Rethinking Image Inpainting via a Mutual Encoder-Decoder with Feature Equalizations" as part of the reproducibility project for the Deep Learning course CS4240 at the TU Delft. Our paper can be found here.
The implementation of our proposed model is a combination of the original MEDFE repository https://github.com/KumapowerLIU/Rethinking-Inpainting-MEDFE and the CycleGAN-and-pix2pix repository https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix of which we just only used the global set-up for the convolutional auto-encoder. Our combined repository with some modifications such as loading the encoder into MEDFE as well as changed ground truth calculations can be found in this repository. For a discussion and motivation together with results and the evaluation of our proposed model, we would like to refer you to our paper here https://github.com/SanderGielisse/MEDFE-CAE/raw/main/MEDFE_CAE_paper.pdf.