PyTorch implementation for "Time Is MattEr: Temporal Self-supervision for Video Transformers" (accepted in ICML 2022)
First, create a conda virtual environment and activate it:
conda create -n TIME python=3.7 -y
source activate TIME
Then, install the following packages:
"torch>=1.9.0"
"torchvision>=0.10.0",
"einops>=0.3",
"yacs>=0.1.6",
"pyyaml>=5.1",
"imageio",
"fvcore",
"timm",
"scikit-learn",
"av",
"matplotlib",
"termcolor>=1.1",
"simplejson",
"tqdm",
"psutil",
"matplotlib",
"scikit-build",
"cmake",
"opencv-python",
"pandas",
"sklearn",
'torchmeta',
"ffmpeg-python",
"tensorboard",
python main.py --cfg configs/timesformer/ssv2/train.yaml
python main.py --cfg configs/motionformer/ssv2/train.yaml
python main.py --cfg configs/xvit/ssv2/train.yaml
You can download the weights of the trained models on Something-Something-V2 (SSv2). All models share the same training details, and they are fine-tuned from the ImageNet-1k pretrained weights.
| backbone | dataset | # of frames | spatial crop | acc@1 | acc@5 | url |
|---|---|---|---|---|---|---|
| TimeSformer | SSv2 | 8 | 224 | 62.1 | 86.4 | model |
| TimeSformer + TIME | SSv2 | 8 | 224 | 63.7 | 87.8 | model |
| Motionformer | SSv2 | 8 | 224 | 63.8 | 88.5 | model |
| Motionformer + TIME | SSv2 | 8 | 224 | 64.7 | 89.3 | model |
| X-ViT | SSv2 | 8 | 224 | 60.1 | 85.2 | model |
| X-ViT + TIME | SSv2 | 8 | 224 | 63.5 | 88.1 | model |
The majority of this work is licensed under CC-NC 4.0 International license. However, portions of the project are available under separate license terms: SlowFast, XViT and pytorch-image-models are licensed under the Apache 2.0 license.
Copyright 2022-present NAVER Corp.
CC BY-NC-4.0 (https://creativecommons.org/licenses/by-nc/4.0/)
Our code base is built partly upon the packages: SlowFast, TimeSformer, Motionformer, X-ViT, and pytorch-image-models by Ross Wightman.
If you use this code for your research, please cite our papers.
@InProceedings{pmlr-v162-yun22a,
title = {Time Is {M}att{E}r: Temporal Self-supervision for Video Transformers},
author = {Yun, Sukmin and Kim, Jaehyung and Han, Dongyoon and Song, Hwanjun and Ha, Jung-Woo and Shin, Jinwoo},
booktitle = {Proceedings of the 39th International Conference on Machine Learning},
pages = {25804--25816},
year = {2022},
editor = {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan},
volume = {162},
series = {Proceedings of Machine Learning Research},
month = {17--23 Jul},
publisher = {PMLR},
pdf = {https://proceedings.mlr.press/v162/yun22a/yun22a.pdf},
url = {https://proceedings.mlr.press/v162/yun22a.html},
abstract = {Understanding temporal dynamics of video is an essential aspect of learning better video representations. Recently, transformer-based architectural designs have been extensively explored for video tasks due to their capability to capture long-term dependency of input sequences. However, we found that these Video Transformers are still biased to learn spatial dynamics rather than temporal ones, and debiasing the spurious correlation is critical for their performance. Based on the observations, we design simple yet effective self-supervised tasks for video models to learn temporal dynamics better. Specifically, for debiasing the spatial bias, our method learns the temporal order of video frames as extra self-supervision and enforces the randomly shuffled frames to have low-confidence outputs. Also, our method learns the temporal flow direction of video tokens among consecutive frames for enhancing the correlation toward temporal dynamics. Under various video action recognition tasks, we demonstrate the effectiveness of our method and its compatibility with state-of-the-art Video Transformers.}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent