futureword123456 / graphfill

GraphFill: Deep Image Inpainting using Graphs

GraphFill

GraphFill: Deep Image Inpainting using Graphs

Project Page

Evaluation

We provide various settings in run.sh

• python main.py config=main.yaml "util_args.predict_only=True" "util_args.eval_mode=True" "data=places_256"

  • predict_only: Setting this flag to False will calculate losses and evaluate metrics. Set True to track performance on validation performance while training.
  • eval_mode: Sets mode for Inference/Training
  • data: change config accordingly at config/data with appropriate paths to training datasets, Validation datasets.

• python data_gen_pickle.py config=main.yaml "data=places_256" "data.train.indir=./Datasets/places365_standard/train"

  • Pyramidal graph generation can be a bottleneck while loading data. Create pickled data for fast loading (Optional).
  • If skipping pickling of data it is recommended to increase num_workers in dataloader kwargs.

• python main.py config=main.yaml "util_args.eval_mode=False" "data=places_256" "data.train.pickle_data=True"

  • Trains GraphFill.

Download trained models from Here

Place downloaded models at as pointed by the key model_load in main.yaml config. Note that shared model contains weights for discriminator, pre-trained model weights for perceptual loss calculation, etc. which are irrelevant in evaluation setting.

Requirements

pytorch_lightning==1.9.0
torch==1.13.1
networkx==2.6
torch_geometric==1.5.0
torch_scatter==2.1.1
torch_sparse==0.6.17 

Code in this repository is highly inspred from: LaMa: Resolution-robust Large Mask Inpainting with Fourier Convolutions. Please follow there Instruction to setup ./models folder, make random masks of sizes medium,thin,thick.

Bib

@inproceedings{verma2024graphfill,
  title={GraphFill: Deep Image Inpainting Using Graphs},
  author={Verma, Shashikant and Sharma, Aman and Sheshadri, Roopa and Raman, Shanmuganathan},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={4996--5006},
  year={2024}
}