Skip to content

Latest commit

 

History

History
103 lines (72 loc) · 3.6 KB

README.org

File metadata and controls

103 lines (72 loc) · 3.6 KB
Raw

SuperSloMo Extended

https://img.shields.io/badge/PyTorch%20-%23EE4C2C.svg https://img.shields.io/badge/code%20style-black-000000.svg

This is a PyTorch implementation of the CVPR 2018 paper, High Quality Estimation of Multiple Intermediate Frames for Video Interpolation. It also includes a few bonus features.

Here’s a sneak peak!

data/teaser.png

If you’d like to watch more examples and comparisons with other state of the art methods, visit the project page!

The nice features in this implementation are:

  1. Multi-GPU training out of the box.
  2. Improved performance on multiple benchmarks as reported in the original paper.
  3. Extensions to the original paper - by using a ConvLSTM/ConvGRU in the bottleneck.

Requirements

This implementation was trained on 4 TitanX GPUs (12GB RAM), with the largest model taking 1 week to train. This code was tested with: python3, PyTorch 1.6, torchvision 0.7, CUDA 9.2, and CUDNN 7.6.3

Setup

Set up a conda environment using configs/conda_env.yml TODO Provide Docker image alternative.

Running the code

All the commands were run using the conda environment specified.

conda activate interp # environment name

Training

PROJECTDIR=<project dir>
LOGDIR=/path/to/logs

cd $PROJECTDIR

...  # edit ssmr.ini to change config settings

# Training
CONFIG=configs/superslomo_original.ini # or superslomo_recurrent.ini
EXPT_NAME=TrainExample01
python scripts/main.py -c $CONFIG --expt=$EXPT_NAME --log=$LOGDIR

Testing Interpolation Results

PROJECTDIR=<project dir>
LOGDIR=/path/to/logs

cd $PROJECTDIR

# Testing interpolation results
CONFIG=configs/superslomo_eval.ini # don't forget to specify the DATASET in the config
EXPT_NAME=TestAdobe
python scripts/evaluate_interpolation_results.py -c $CONFIG --log=$LOGDIR --expt=$EXPT_NAME

Testing Optical Flow Results

PROJECTDIR=<project dir>
LOGDIR=/path/to/logs

cd $PROJECTDIR

# Testing optical flow results
CONFIG=configs/superslomo_eval.ini
EXPT_NAME=TestOpticalFlow
python scripts/evaluate_optical_flow_results.py -c $CONFIG --log=$LOGDIR

Visualizing Interpolation Results

# Visualizing the interpolation results

IMGDIR=Adobe240fps/Clips/clip_00074
OUTDIR=Results/
echo "Processing: "$IMGDIR
CONFIG=configs/superslomo_eval.ini

python scripts/visualize_interpolation.py -c $CONFIG --expt=$EXPT_NAME --log=$LOGDIR --input_dir=$IMGDIR --output_dir=$OUTDIR --img_type=png --upsample_rate=8 --is_fps_240

Performance Benchmarks

Please refer to the tech report, and experiments/Notes.org for comparisons with the original paper, and other approaches.

Acknowledgements

Many thanks to the author of the original paper (Huaizu Jiang) for clarifying several key implementation details. Thanks also to Prof. Erik Learned-Miller and Prof. Liangliang Cao (both at UMass Amherst) for their support and advice during this project.

Other Implementations

This paper is quite popular, and has been previously implemented here, here, and here.

LICENSE

TODO

Contact

Please email me at sreenivas.vrao1@gmail.com if you have questions.