Video Motion Magnification

This repository implements T. Oh, R. Jaroensri, C. Kim, M. Elgharib, F. Durand, W. Freeman, W. Matusik "Learning-based Video Motion Magnification" arXiv preprint arXiv:1804.02684 (2018) in PyTorch.

Please visit the homepage of the research publication above to see the authors' excellent paper, high-quality TensorFlow code, and a ready-to-use training dataset. The authors did an amazing job, which made it easy for me to create this PyTorch implementation. 🙏🏽

"Motion magnification" means you pick up and amplify small motions so they're easier to see. It's like a microscope for motion.

Samples

👉 Click here to see sample videos from this implementation 👈

You can download the videos here.

Set up

This implementation requires Python 3 and PyTorch. Install dependencies with pip:

pip install -r requirements.txt

Run (examples)

# Get CLI help.
python main.py -h

# Amplify video 5x.
python main.py amplify data/models/20191204-b4-r0.1-lr0.0001-05.pt data/examples/baby.mp4 --amplification=5

# Amplify 7x and render side-by-side comparison video.
python main.py demo data/models/20191204-b4-r0.1-lr0.0001-05.pt data/examples/baby.mp4 baby-demo.mp4 7

# Make video collage with input, 2x and 4x amplified side by side.
python main.py demo data/models/20191204-b4-r0.1-lr0.0001-05.pt data/examples/baby.mp4 baby-demo.mp4 2 4

Training

I've included a pre-trained model so you can try it out of the box, but you can also train from scratch if you wish.

First, download and extract the training dataset published by the authors. It's ~84GB in total compressed, and slighlty larger deflated. Make sure you have enough disk space.

Example train commands:

# Get CLI help and adjust training parameters with flags.
python main.py train -h

# Train from scratch - this will save a model on disk after each epoch.
python main.py train data/train data/models --num-epochs=10

# Resume training from previously saved snapshot and skip 3 epochs.
python main.py train data/train data/models --num-epochs=10 --skip-epochs=3 --load-model-path=data/models/20191203-b10-r0.1-lr0.0001-02.pt

This implementation takes ~2 days to arrive at a decent loss on a Tesla P100 GPU. Let me know if you speed it up.