Home

Awesome

FuseFormer: Fusing Fine-Grained Information in Transformers for Video Inpainting

By Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, Hongsheng Li.

This repo is the official Pytorch implementation of FuseFormer: Fusing Fine-Grained Information in Transformers for Video Inpainting.

Introduction

<img src='imgs/pipeline.png' width="900px">

Usage

Prerequisites

Install

git clone https://github.com/ruiliu-ai/FuseFormer.git
cd FuseFormer
pip install -r requirements.txt

Training

Dataset preparation

Download datasets (YouTube-VOS and DAVIS) into the data folder.

mkdir data

Note: We use YouTube Video Object Segmentation dataset 2019 version.

Training script

python train.py -c configs/youtube-vos.json

Test

Download pre-trained model into checkpoints folder.

mkdir checkpoints

Test script

python test.py -c checkpoints/fuseformer.pth -v data/DAVIS/JPEGImages/blackswan -m data/DAVIS/Annotations/blackswan

Evaluation

You can follow free-form mask generation scheme for synthesizing random masks.

Or just download our prepared stationary masks and unzip it to data folder.

mv random_mask_stationary_w432_h240 data/
mv random_mask_stationary_youtube_w432_h240 data/

Then you need to download pre-trained model for evaluating VFID.

mv i3d_rgb_imagenet.pt checkpoints/

Evaluation script

python evaluate.py --model fuseformer --ckpt checkpoints/fuseformer.pth --dataset davis --width 432 --height 240
python evaluate.py --model fuseformer --ckpt checkpoints/fuseformer.pth --dataset youtubevos --width 432 --height 240

For evaluating warping error, please refer to https://github.com/phoenix104104/fast_blind_video_consistency

Citing FuseFormer

If you find FuseFormer useful in your research, please consider citing:

@InProceedings{Liu_2021_FuseFormer,
  title={FuseFormer: Fusing Fine-Grained Information in Transformers for Video Inpainting},
  author={Liu, Rui and Deng, Hanming and Huang, Yangyi and Shi, Xiaoyu and Lu, Lewei and Sun, Wenxiu and Wang, Xiaogang and Dai, Jifeng and Li, Hongsheng},
  booktitle = {International Conference on Computer Vision (ICCV)},
  year={2021}
}

Acknowledement

This code borrows heavily from the video inpainting framework spatial-temporal transformer net.