Progressive Temporal Feature Alignment Network for Video Inpainting

This work is accepted in CVPR2021 as Poster. It proposed a new video inpainting approach that combines temporal convolution as well as optical flow approach.

Noted: This code is currently a beta version. Not gurantee to be fully correct.

Update

Optical Flow Davis | Optical Flow FVI | Mask Davis | Mask FVI | Checkpoint

Installation

torch==1.7.0
torchvision==0.8.1

Dataset

For FVI dataset, please refer to https://github.com/amjltc295/Free-Form-Video-Inpainting. For DAVIS dataset, please refer to https://davischallenge.org/.

File Structure

TSAM
└── data
    ├── checkpoints
    ├── model_weights
    ├── results
    ├── FVI
    ├── DAVIS    
    └── runs
└── code
    └── master
        └── TSAM
            └── ...

Prepare pretrained weights for training

Pretrained weights: download all the pretrained weights and put it under TSAM/data/model_weights

Model Name
TSM_imagenet_resent50_gated.pth	weight
TSM_imagenet_resent50.pth	weight

Training

FVI TSM moving object/curve masks:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train.py --config config/config_pretrain.json --dataset_config dataset_configs/FVI_all_masks.json
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train.py --config config/config_finetune.json --dataset_config dataset_configs/FVI_all_masks.json

Testing

Change the train.py in training scripts to test.py, and add -p /pth/to/ckpt to the end.

DAVIS TSAM object removal:

CUDA_VISIBLE_DEVICES=0 python3 test.py --config config/config_finetune_davis.json --dataset_config dataset_configs/DAVIS_removal.json -p /pth/to/ckpt

Citation

@inproceedings{zou2020progressive,
  title={Progressive Temporal Feature Alignment Network for Video Inpainting},
  author={Xueyan Zou and Linjie Yang and Ding Liu and Yong Jae Lee},
  booktitle={CVPR},
  year={2021}
}

Acknowledgement

Part of the code is borrow from https://github.com/amjltc295/Free-Form-Video-Inpainting and https://github.com/researchmm/STTN. Thanks for their great works!