Skip to content

lukas-blecher/Deep-Flow-Guided-Video-Inpainting

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

50 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Deep Flow-Guided Video Inpainting

CVPR 2019 Paper | Project Page | YouTube | BibeTex

Install & Requirements

The code has been tested on pytorch=1.3.0 and python3.6. Please refer to requirements.txt for detailed information.

Alternatively, you can run it with the provided Docker image.

To Install python packages

pip install -r requirements.txt

To Install liteflownet modules

The correlation layer for LiteFlowNet is implemented in CUDA using CuPy. Install it using pip install cupy or install one of the provided binaries (listed here).

Componets

There exist three components in this repo:

Usage

  • To use our video inpainting tool for object removing, we recommend that the frames should be put into xxx/video_name/frames and the mask of each frame should be put into xxx/video_name/masks. And please download the resources of the demo and model weights from here. An example demo containing frames and masks has been put into the demo and running the following command will get the result:
python tools/video_inpaint.py --frame_dir ./demo/frames --MASK_ROOT ./demo/masks --img_size 512 832 --LiteFlowNet --DFC --ResNet101 --Propagation 

We provide the original model weight used in our movie demo which use ResNet101 as backbone and other related weights pls download from here. Weights for LiteFlowNet are hosted by sniklaus: default, kitti, sintel. Please refer to tools for detailed use and training settings.

  • For fixed region inpainting, we provide the model weights of refined stages in DAVIS. Please download the lady-running resources link and model weights link. The following command can help you to get the result:
CUDA_VISIBLE_DEVICES=0 python tools/video_inpaint.py --frame_dir ./demo/lady-running/frames \
--MASK_ROOT ./demo/lady-running/mask_bbox.png \
--img_size 448 896 --DFC --LiteFlowNet --Propagation \
--PRETRAINED_MODEL_1 ./pretrained_models/resnet50_stage1.pth \
--PRETRAINED_MODEL_2 ./pretrained_models/DAVIS_model/davis_stage2.pth \
--PRETRAINED_MODEL_3 ./pretrained_models/DAVIS_model/davis_stage3.pth \
--MS --th_warp 3 --FIX_MASK

You can just change the **th_warp** param for getting better results in your video.
  • To extract flow for videos:
python tools/infer_liteflownet.py --frame_dir xxx/video_name/frames
  • To use the Deepfillv1-Pytorch model for image inpainting,
python tools/frame_inpaint.py --test_img xxx.png --test_mask xxx.png --image_shape 512 512

Update

  • More results can be found and downloaded here.

  • Support for PyTorch>1.0: Sorry for the late update and the pre-release verison for supporting PyTorch>1.0 has been integrated into our new v1.1 branch.

  • The frames and masks of our movie demo have been put into Google Drive.

  • The weights of DAVIS's refined stages have been released and you can download from here. Please refer to Usage for using the Multi-Scale models.

FAQ

  • Errors when running install_scripts.sh if you meet some problem about gcc when compiling, pls check if the following commands will help:
export CXXFLAGS="-std=c++11"
export CFLAGS="-std=c99"

Citation

@InProceedings{Xu_2019_CVPR,
author = {Xu, Rui and Li, Xiaoxiao and Zhou, Bolei and Loy, Chen Change},
title = {Deep Flow-Guided Video Inpainting},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}

About

pytorch implementation for "Deep Flow-Guided Video Inpainting"(CVPR'19)

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 99.6%
  • Dockerfile 0.4%