Learning Blind Video Temporal Consistency (ECCV 2018)
Switch branches/tags
Nothing to show
Clone or download
Latest commit 1d1a964 Dec 5, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
data test pretrained Jul 31, 2018
lists test pretrained Jul 31, 2018
networks update FlowNet2 Dec 6, 2018
pretrained_models update download_models Aug 1, 2018
.gitignore add test_pretrained Jul 31, 2018
README.md Update README.md Aug 3, 2018
batch_evaluate.py fix inputs bux Sep 22, 2018
batch_test.py fix inputs bux Sep 22, 2018
compute_flow_occlusion.py init Jul 30, 2018
datasets.py fix inputs bux Sep 22, 2018
evaluate_LPIPS.py init Jul 30, 2018
evaluate_WarpError.py init Jul 30, 2018
install.sh init Jul 30, 2018
teaser_small.gif Add files via upload Jul 31, 2018
test.py fix inputs bux Sep 22, 2018
test_pretrained.py update test_pretrained Dec 4, 2018
train.py fix inputs bux Sep 22, 2018
utils.py init Jul 30, 2018


Learning Blind Video Temporal Consistency

Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang

European Conference on Computer Vision (ECCV), 2018

[Project page][Paper]

Table of Contents

  1. Introduction
  2. Requirements and Dependencies
  3. Installation
  4. Dataset
  5. Apply Pre-trained Models
  6. Training and Testing
  7. Evaluation
  8. Image Processing Algorithms
  9. Citation


Our method takes the original unprocessed and per-frame processed videos as inputs to produce a temporally consistent video. Our approach is agnostic to specific image processing algorithms applied on the original video.

Requirements and dependencies

Our code is tested on Ubuntu 16.04 with cuda 9.0 and cudnn 7.0.


Download repository:

git clone https://github.com/phoenix104104/fast_blind_video_consistency.git

Compile FlowNet2 dependencies (correlation, resample, and channel norm layers):



Download our training and testing datasets:

cd data
./download_data.sh [train | test | all]
cd ..

For example, download training data only:

./download_data.sh train

Download both training and testing data:

./download_data.sh all

You can also download the results of [Bonneel et al. 2015] and our approach:

./download_data.sh results

Apply pre-trained models

Download pretrained models (including FlowNet2 and our model):

cd pretrained_models
cd ..

Test pre-trained model:

python test_pretrained.py -dataset DAVIS -task WCT/wave

The output frames are saved in data/test/ECCV18/WCT/wave/DAVIS.

Training and testing

Train a new model:

python train.py -datasets_tasks W3_D1_C1_I1

We have specified all the default parameters in train.py. lists/train_tasks_W3_D1_C1_I1.txt specifies the dataset-task pairs for training.

Test a model:

python test.py -method MODEL_NAME -epoch N -dataset DAVIS -task WCT/wave

Check the checkpoint folder for the MODEL_NAME. The output frames are saved in data/test/MODEL_NAME/epoch_N/WCT/wave/DAVIS.

You can also generate results for multiple tasks using the following script:

python batch_test.py -method output/MODEL_NAME/epoch_N

which will test all the tasks in lists/test_tasks.txt.


Temporal Warping Error

To compute the temporal warping error, we first need to generate optical flow and occlusion masks:

python compute_flow_occlusion.py -dataset DAVIS -phase test

The flow will be stored in data/test/fw_flow/DAVIS. The occlusion masks will be stored in data/test/fw_occlusion/DAVIS.

Then, run the evaluation script:

python evaluate_WarpError.py -method output/MODEL_NAME/epoch_N -task WCT/wave


Download LPIPS repository and change LPIPS_dir in evalate_LPIPS.py if necesary (default path is ../LPIPS).

Run the evaluation script:

python evaluate_LPIPS.py -method output/MODEL_NAME/epoch_N -task WCT/wave

Batch evaluation

You can evaluate multiple tasks using the following script:

python batch_evaluate.py -method output/MODEL_NAME/epoch_N -metric LPIPS
python batch_evaluate.py -method output/MODEL_NAME/epoch_N -metric WarpError

which will evaluate all the tasks in lists/test_tasks.txt.

Test on new videos

To test our model on new videos or applications, please follow the folder structure in ./data.

Given a video, we extract frames named as %05d.jpg and save frames in data/test/input/DATASET/VIDEO.

The per-frame processed video is stored in data/test/processed/TASK/DATASET/VIDEO, where TASK is the image processing algorithm applied on the original video.

Image Processing Algorithms

We use the following algorithms to obtain per-frame processed results:

Style transfer

Image Enhancement

Intrinsic Image Decomposition

Image-to-Image Translation



If you find the code and datasets useful in your research, please cite:

    author    = {Lai, Wei-Sheng and Huang, Jia-Bin and Wang, Oliver and Shechtman, Eli and Yumer, Ersin and Yang, Ming-Hsuan}, 
    title     = {Learning Blind Video Temporal Consistency}, 
    booktitle = {European Conference on Computer Vision},
    year      = {2018}