jbelof / super-slomo

PyTorch implementation of Super SloMo by Jiang et al.

Super-SloMo

PyTorch implementation of "Super SloMo: High Quality Estimation of Multiple Intermediate Frames for Video Interpolation" by Jiang H., Sun D., Jampani V., Yang M., Learned-Miller E. and Kautz J. [Project] [Paper]

Check out our paper "Deep Slow Motion Video Reconstruction with Hybrid Imaging System" published in TPAMI.

Results

Results on UCF101 dataset using the evaluation script provided by paper's author. The get_results_bug_fixed.sh script was used. It uses motions masks when calculating PSNR, SSIM and IE.

Method	PSNR	SSIM	IE
DVF	29.37	0.861	16.37
SepConv - L_1	30.18	0.875	15.54
SepConv - L_F	30.03	0.869	15.78
SuperSloMo_Adobe240fps	29.80	0.870	15.68
pretrained mine	29.77	0.874	15.58
SuperSloMo	30.22	0.880	15.18

Prerequisites

This codebase was developed and tested with pytorch 0.4.1 and CUDA 9.2 and Python 3.6. Install:

• PyTorch

For GPU, run

conda install pytorch=0.4.1 cuda92 torchvision==0.2.0 -c pytorch

For CPU, run

conda install pytorch-cpu=0.4.1 torchvision-cpu==0.2.0 cpuonly -c pytorch

• TensorboardX for training visualization
• tensorflow for tensorboard
• matplotlib for training graph in notebook.
• tqdm for progress bar in video_to_slomo.py
• numpy

Training

Preparing training data

In order to train the model using the provided code, the data needs to be formatted in a certain manner. The create_dataset.py script uses ffmpeg to extract frames from videos.

Adobe240fps

For adobe240fps, download the dataset, unzip it and then run the following command

python data\create_dataset.py --ffmpeg_dir path\to\folder\containing\ffmpeg --videos_folder path\to\adobe240fps\videoFolder --dataset_folder path\to\dataset --dataset adobe240fps

Custom

For custom dataset, run the following command

python data\create_dataset.py --ffmpeg_dir path\to\folder\containing\ffmpeg --videos_folder path\to\adobe240fps\videoFolder --dataset_folder path\to\dataset

The default train-test split is 90-10. You can change that using command line argument --train_test_split.

Run the following commmand for help / more info

python data\create_dataset.py --h

Training

In the train.ipynb, set the parameters (dataset path, checkpoint directory, etc.) and run all the cells.

or to train from terminal, run:

python train.py --dataset_root path\to\dataset --checkpoint_dir path\to\save\checkpoints

Run the following commmand for help / more options like continue from checkpoint, progress frequency etc.

python train.py --h

Tensorboard

To get visualization of the training, you can run tensorboard from the project directory using the command:

tensorboard --logdir log --port 6007

and then go to https://localhost:6007.

Evaluation

Pretrained model

You can download the pretrained model trained on adobe240fps dataset here.

Video Converter

You can convert any video to a slomo or high fps video (or both) using video_to_slomo.py. Use the command

# Windows
python video_to_slomo.py --ffmpeg path\to\folder\containing\ffmpeg --video path\to\video.mp4 --sf N --checkpoint path\to\checkpoint.ckpt --fps M --output path\to\output.mkv

# Linux
python video_to_slomo.py --video path\to\video.mp4 --sf N --checkpoint path\to\checkpoint.ckpt --fps M --output path\to\output.mkv

If you want to convert a video from 30fps to 90fps set fps to 90 and sf to 3 (to get 3x frames than the original video).

Run the following commmand for help / more info

python video_to_slomo.py --h

You can also use eval.py if you do not want to use ffmpeg. You will instead need to install opencv-python using pip for video IO. A sample usage would be:

python eval.py data/input.mp4 --checkpoint=data/SuperSloMo.ckpt --output=data/output.mp4 --scale=4

Use python eval.py --help for more details

More info TBA

References:

Parts of the code is based on TheFairBear/Super-SlowMo