Intro

The official code for the following paper :

Boundary-aware Image Inpainting with Multiple Auxiliary Cues, NTIRE2022

Download paper here

Prerequisites

1.Clone the repository

git clone https://github.com/rain58/Boundary-aware-Image-Inpainting.git  
cd Boundary-aware-Image-Inpainting

2.To create Python environment:

pip install -r requirements.txt

We use Places2 and Paris Street-View.Please download the datasets from their official website. Arter downloading, make flist files:

mkdir datasets
python ./scripts/flist.py --path path_to_rgb_train_set --output ./datasets/rgb_train.flist

Estimate depth image from RGB Images datasets by using Dense Depth. The procedure is as follows.

Fine-tune the pre-trained Dense Depth model by using DIODE. We only use outdoor images of DIODE dataset. Please download from here
Estimate the depth image from RGB Images.
make flist files:

python ./scripts/flist.py --path path_to_depth_train_set --output ./datasets/depth_train.flist

We use mask datasets provided by Liu et al.. You can download datasets from here
Arter downloading, make flist files:

python ./scripts/flist.py --path path_to_mask_train_set --output ./datasets/mask_train.flist

To train the model, create a config.yaml file similar to the example config file and copy it under your checkpoints directory.

Our model is trained in three stages:

training edge model or download the edge model from EdgeConnect
training depth model
training the inpaint model
To train the model, change the "model" option number in train.sh.
For examples, Edge:1, Depth:5, Inpaint:6. Then, run

sh train.sh

Download the pre-trained models from the following links and copy them under ./checkpoints
our pre-trained model : Paris Street-View, Places

To test our model, change the dataset path in test.sh and run

sh test.sh

This work is based on the EdgeConnect. We very appreciate their hard work and great research.