Learning Temporally Invariant and Localizable Features via Data Augmentation for Video Recognition (ECCVW 2020)

Taeoh Kim^*1, Hyeongmin Lee^*1, MyeongAh Cho^*1, HoSeong Lee², Dong Heon Cho², Sangyoun Lee^1
* indicates equal contribution
¹ Yonsei University
² Cognex Deep Learning Lab

Introduction

Official PyTorch implementation of our paper which has been accepted to 1st Visual Inductive Priors for Data-Efficient Deep Learning Workshop at ECCV 2020 as Oral Presentation.

Paper Link

Environment

• PyTorch 1.4.0
• opencv-python
• PIL (pillow)
• tqdm

Train

Prepare training data (Modified UCF-101)

1. Download UCF-101 training data from UCF-101 dataset.

2. Modify training data using VIPriors Challenge Split.

How to train

1. Run train.py with following command. (SlowFast-50 Baseline)

   python train.py --dir_data [DB] --out_dir [ExpName] --is_validate

2. Train on RandAugment-T

   python train.py --dir_data [DB] --out_dir [ExpName] --is_validate --rand_augmentation --aug_mode randaug --randaug_n [N] --randaug_m [M]

3. Train on Mixing Data Augmentations (cutmix, framecutmix, cubecutmix, mixup, fademixup, cutmixup, framecutmixup, cubecutmixup)

   python train.py --dir_data [DB] --out_dir [ExpName] --is_validate --mix_type [Mix_type]

How to test

1. Run train.py with following command.

   python train.py --dir_data [DB] --out_dir [ExpName] --load [ckpt] --test_only --is_validate

Citation

If you find the code helpful in your resarch or work, please cite the following paper.

@inproceedings{kim2020learning,
    title={Learning Temporally Invariant and Localizable Features via Data Augmentation for Video Recognition},
    author={Kim, Taeoh and Lee, Hyeongmin and Cho, MyeongAh and Lee, Ho Seong and Cho, Dong Heon and Lee, Sangyoun},
    booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
    year={2020}
}