baselines

Image inpainting Baselines

This repo is created to provide EASY comparison with SOME Image Inpainting methods. Feel free to create a pull request to add more.

Also, keep in mind there are other methods that leverage on deep priors for the inpainting task such as Deep Image Prior, PsP, and more novel Diffusion Models such as SDEdit or RePaint.

Run Baseline

Each folder has a run.sh that installs all the dependencies** and evaluate each method (from their original implementations and their original pretrained weights) according to three variables as follows:

input_dir=/path/to/inputdir output_dir=/path/to/outputdir DATASET=celeba sh run.sh

• input_dir: folder with both images and masks in a 1:1 ratio. Masks should have the name mask in the filename. It only supports png images.
• output_dir: it will contain the generated images with the same name as the input images.
• DATASET: depending on the selected method, it could be celeba, imagenet, or places.

**Modify run.sh to your needs.

Test Tar files

Sometimes the test evaluation is given by several tar files with different masks and different images, as follows:

tar_folder
|_ images_gt_a.tar
|_ masks_gt_a.tar
|_ images_gt_b.tar
|_ masks_gt_b.tar
.
.
.

In such case, you should run test.py with the method as argument (eg., python test.py --method ict), and it will create a folder with the name of the method, as follows (for instance ICT method):

tar_folder
|_ images_ict256_gt_a.tar
|_ masks_ict256_gt_a.tar
|_ images_c256_gt_b.tar
|_ masks_c256_gt_b.tar
|_ ict
  |_ images_ict256_gt_a.tar
  |_ images_c256_gt_b.tar
.
.
.

It will dynamically change the dataset if the tar filename has c256 for celeba and inet256 for imagenet. Modify test.py to your needs.

Requirements

Each method requirement is installed internally. The only additional requirement is gshell, which is used to download the pretrained weights from Google Drive. It might ask you for your authorization in your google account.