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

Introduction

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

• Pytorch 0.4.0 (FlowNet2 dependencies cannot be compiled on Pytorch 0.4.1)
• TensorboardX
• FlowNet2-Pytorch (Code and model are already included in this repository)
• LPIPS (for evaluation)

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

Installation

Download repository:

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

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

./install.sh

Dataset

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
./download_models.sh
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.

Evaluation

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

LPIPS

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

• WCT: Universal Style Transfer via Feature Transforms, NIPS 2017
• Fast Neural Style Transfer: Perceptual Losses for Real-Time Style Transfer and Super-Resolution, ECCV 2016

Image Enhancement

• DBL: Deep Bilateral Learning for Real-Time Image Enhancement, Siggraph 2017

Intrinsic Image Decomposition

• Intrinsic Images in the Wild, Siggraph 2014

Image-to-Image Translation

• CycleGAN: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks, ICCV 2017

Colorization

• Colorful Image Colorization, ECCV 2016
• Let there be Color!: Joint End-to-end Learning of Global and Local Image Priors for Automatic Image Colorization with Simultaneous Classification, Siggraph 2016

Citation

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

@inproceedings{Lai-ECCV-2018,
    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}
}