View Code? Open in Web Editor NEW

Official PyTorch Implementation of MotionSqueeze, ECCV 2020

License: BSD 2-Clause "Simplified" License

Python 81.50% Shell 7.27% Jupyter Notebook 11.24%

motionsqueeze's Introduction

MotionSqueeze: Neural Motion Feature Learning for Video Understanding

This is the implementation of the paper "MotionSqueeze: Neural Motion Feature Learning for Video Understanding" by H.Kwon, M.Kim, S.Kwak, and M.Cho. For more information, checkout the project website and the paper on arXiv.

Environment:

Cuda: 9.0
gcc: 7.3.0
Python 3.6.8
PyTorch 1.0.1
TorchVison: 0.2.2
Spatial Correlation Sampler (https://github.com/ClementPinard/Pytorch-Correlation-extension.git)
Others: environment

Clone this repo

git clone https://github.com/arunos728/MotionSqueeze.git

Anaconda environment setting

cd MotionSqueeze
conda env create -f environment.yml
conda activate MS

Installing Correlation sampler

cd Pytorch-Correlation-extension
python setup.py install

Please check this repo for the detailed instructions.

Running

For training TSM or MSNet on Something-v1, use the following command:

    ./scripts/train_TSM_Something_v1.sh local

For training TSM or MSNet on Kinetics, use the following command:

    ./scripts/train_TSM_Kinetics.sh local

For testing your trained model on Something-v1, use the following command:

    ./scripts/test_TSM_Something_v1.sh local

For testing your trained model on Kinetics, use the following command:

    ./scripts/test_TSM_Kinetics.sh local

Citation

If you use this code or ideas from the paper for your research, please cite our paper:

@inproceedings{kwon2020motionsqueeze,
title={MotionSqueeze: Neural Motion Feature Learning for Video Understanding},	  
author={Heeseung Kwon and Manjin Kim and Suha Kwak and Minsu Cho},     
booktitle={ECCV},
year={2020}
}