darengking / Hifill-tensorflow

Tensorflow reimplementation of Hifill

tensorflow reimplementation of HiFill (CVPR 2020 Oral Paper)

Contextual Residual Aggregation for Ultra High-Resolution Image Inpainting

If the codes helps you in reserach, please cite the following paper:

@misc{yi2020contextual,
    title={Contextual Residual Aggregation for Ultra High-Resolution Image Inpainting},
    author={Zili Yi and Qiang Tang and Shekoofeh Azizi and Daesik Jang and Zhan Xu},
    year={2020},
    eprint={2005.09704},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Youtube video

how to train

get dependancies installed on the conda environment, see requirements.txt for details. The following package are required:

• tensorflow-gpu
• opencv
• scipy
• pyyaml
• neuralgym
• easydict

get data prepared as in ./data/examples/

specify the training data path in config.yaml

run the following script to start training. 200000~300000 steps would be enough for good converngence

python train.py

how to test

python test.py --image_dir='./data/test/images' --mask_dir='./data/test/masks' --output_dir='outputs' --checkpoint_dir='./model_logs/places2' --input_size=512 --times=8

Exemplar Experimental results: