DAIN (Depth-Aware Video Frame Interpolation)

Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang

IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, CVPR 2019

THIS WORK IS BUILT UPON OUR PREVIOUS WORK CALLED MEMC-NET, IN WHICH WE PROPOSE THE ADAPTIVE WARPING LAYER. PLEASE ALSO CONSIDER CITING MEMC-NET.

Introduction
Citation
Requirements and Dependencies
Installation
Testing Pre-trained Models
Downloading Results
Slow-motion Generation
Training New Models

Introduction

We propose the Depth-Aware video frame INterpolation (DAIN) model to explicitly detect the occlusion by exploring the depth cue. We develop a depth-aware flow projection layer to synthesize intermediate flows that preferably sample closer objects than farther ones. Our method achieves state-of-the-art performance on the Middlebury dataset. We provide videos here.

Citation

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

@inproceedings{DAIN,
    author    = {Bao, Wenbo and Lai, Wei-Sheng and Ma, Chao and Zhang, Xiaoyun and Gao, Zhiyong and Yang, Ming-Hsuan}, 
    title     = {Depth-Aware Video Frame Interpolation}, 
    booktitle = {IEEE Conferene on Computer Vision and Pattern Recognition},
    year      = {2019}
}

Requirements and Dependencies

Ubuntu (We test with Ubuntu = 16.04.5 LTS)
Python (We test with Python = 3.6.8 in Anaconda3 = 4.1.1)
Cuda & Cudnn (We test with Cuda = 9.0 and Cudnn = 7.0)
PyTorch (The customized depth-aware flow projection and other layers require ATen API in PyTorch = 1.0.0)
GCC (Compiling PyTorch 1.0.0 extension files (.c/.cu) requires gcc = 4.9.1 and nvcc = 9.0 compilers)
NVIDIA GPU (We use Titan X (Pascal) with compute = 6.1, but we support compute_50/52/60/61 devices, should you have devices with higher compute capability, please revise this)

Installation

Download repository:

$ git clone https://github.com/baowenbo/DAIN.git

Before building Pytorch extensions, be sure you have pytorch >= 1.0.0:

$ python -c "import torch; print(torch.__version__)"

Generate our PyTorch extensions:

$ cd DAIN
$ cd my_package 
$ ./build.sh

Generate the Correlation package required by PWCNet:

$ cd ../PWCNet/correlation_package_pytorch1_0
$ ./build.sh

Testing Pre-trained Models

Make model weights dir and Middlebury dataset dir:

$ cd DAIN
$ mkdir model_weights
$ mkdir MiddleBurySet

Download pretrained models,

$ cd model_weights
$ wget http://vllab1.ucmerced.edu/~wenbobao/DAIN/best.pth

and Middlebury dataset:

$ cd ../MiddleBurySet
$ wget http://vision.middlebury.edu/flow/data/comp/zip/other-color-allframes.zip
$ unzip other-color-allframes.zip
$ wget http://vision.middlebury.edu/flow/data/comp/zip/other-gt-interp.zip
$ unzip other-gt-interp.zip
$ cd ..

We are good to go by:

$ CUDA_VISIBLE_DEVICES=0 python demo_MiddleBury.py

The interpolated results are under MiddleBurySet/other-result-author/[random number]/, where the random number is used to distinguish different runnings.

Downloading Results

Our DAIN model achieves the state-of-the-art performance on the UCF101, Vimeo90K, and Middlebury (eval and other). Dowload our interpolated results with:

$ wget http://vllab1.ucmerced.edu/~wenbobao/DAIN/UCF101_DAIN.zip
$ wget http://vllab1.ucmerced.edu/~wenbobao/DAIN/Vimeo90K_interp_DAIN.zip
$ wget http://vllab1.ucmerced.edu/~wenbobao/DAIN/Middlebury_eval_DAIN.zip
$ wget http://vllab1.ucmerced.edu/~wenbobao/DAIN/Middlebury_other_DAIN.zip

Slow-motion Generation

Our model is fully capable of generating slow-motion effect with minor modification on the network architecture. Run the following code by specifying time_step = 0.25 to generate x4 slow-motion effect:

$ CUDA_VISIBLE_DEVICES=0 python demo_MiddleBury_slowmotion.py --netName DAIN_slowmotion --time_step 0.25

or set time_step to 0.125 or 0.1 as follows

$ CUDA_VISIBLE_DEVICES=0 python demo_MiddleBury_slowmotion.py --netName DAIN_slowmotion --time_step 0.125
$ CUDA_VISIBLE_DEVICES=0 python demo_MiddleBury_slowmotion.py --netName DAIN_slowmotion --time_step 0.1

to generate x8 and x10 slow-motion respectively. Or if you would like to have x100 slow-motion for a little fun.

$ CUDA_VISIBLE_DEVICES=0 python demo_MiddleBury_slowmotion.py --netName DAIN_slowmotion --time_step 0.01

You may also want to create gif animations by:

$ cd MiddleBurySet/other-result-author/[random number]/Beanbags
$ convert -delay 1 *.png -loop 0 Beanbags.gif //1*10ms delay

Have fun and enjoy yourself!

Training New Models

Download the Vimeo90K triplet dataset for video frame interpolation task, also see here by Xue et al., IJCV19.

$ cd DAIN
$ mkdir /path/to/your/dataset & cd /path/to/your/dataset 
$ wget http://data.csail.mit.edu/tofu/dataset/vimeo_triplet.zip
$ unzip vimeo_triplet.zip
$ rm vimeo_triplet.zip

Download the pretrained MegaDepth and PWCNet models

$ cd MegaDepth/checkpoints/test_local
$ wget http://vllab1.ucmerced.edu/~wenbobao/DAIN/best_generalization_net_G.pth
$ cd ../../../PWCNet
$ wget http://vllab1.ucmerced.edu/~wenbobao/DAIN/pwc_net.pth.tar
$ cd  ..

Run the training script:

$ CUDA_VISIBLE_DEVICES=0 python train.py --datasetPath /path/to/your/dataset --batch_size 1 --save_which 1 --lr 0.0001 --rectify_lr 0.0001 --flow_lr_coe 0.01 --occ_lr_coe 0.0 --filter_lr_coe 1.0 --ctx_lr_coe 1.0 --alpha 0.0 1.0 --patience 4 --factor 0.2

The optimized models will be saved to the model_weights/[random number] directory, where [random number] is generated for different runs.

Replace the pre-trained model_weights/best.pth model with the newly trained model_weights/[random number]/best.pth model. Then test the new model by executing:

$ CUDA_VISIBLE_DEVICES=0 python demo_MiddleBury.py

Contact

Wenbo Bao; Wei-Sheng (Jason) Lai

License

See MIT License

gibsonhan/DAIN