# Spatiotemporal Contrastive Video Representation Learning

[![Paper](http://img.shields.io/badge/Paper-arXiv.2008.03800-B3181B?logo=arXiv)](https://arxiv.org/abs/2008.03800)

This repository is the official TF2 implementation of [Spatiotemporal Contrastive Video Representation Learning](https://arxiv.org/abs/2008.03800).

<p align="left">
  <img src="https://storage.googleapis.com/tf_model_garden/vision/cvrl/artifacts/cvrl_overview.png" height=350>
</p>

## Description

We present a self-supervised Contrastive Video Representation Learning (CVRL)
method to learn spatiotemporal visual representations from unlabeled videos. Our
representations are learned using a contrastive loss, where two augmented clips
from the same short video are pulled together in the embedding space, while
clips from different videos are pushed away. CVRL significantly closes the gap
between unsupervised and supervised video representation learning.

We release the code and pre-trained models.

More pre-trained model checkpoints and a detailed instruction about the code
will be updated.


## Experimental Results

### Kinetics-600 top-1 linear classification accuracy

<p align="left">
  <img src="https://storage.googleapis.com/tf_model_garden/vision/cvrl/artifacts/cvrl_results.png" height=350>
</p>


## Pre-trained Model Checkpoints

We provide model checkpoints pre-trained on unlabeled RGB videos from
Kinetics-400 and Kinetics-600. All models are trained scratch with random
initialization.

We also provide a baseline model checkpoint of "ImageNet inflated" we used in
the paper. The model has the same architecture as 3D-ResNet-50 (R3D-50), with
model weights inflated from a 2D ResNet-50 pre-trained on ImageNet.

| Model | Parameters | Dataset | Epochs | K400 Linear Eval. | K600 Linear Eval. | Checkpoint |
| :--------------:  | :----: | :--: | :--: |:-----------: | :----------: | :----------: |
| R3D-50 (1x) | 31.7M  | ImageNet | - | 53.5% | 54.7% | [ckpt (127 MB)](https://storage.googleapis.com/tf_model_garden/vision/cvrl/imagenet.tar.gz) |
| R3D-50 (1x) | 31.7M  | Kinetics-400 | 200 | 63.8% | - | [ckpt (127 MB)](https://storage.googleapis.com/tf_model_garden/vision/cvrl/r3d_1x_k400_200ep.tar.gz) |
| R3D-50 (1x) | 31.7M  | Kinetics-400 | 800 | 66.1% | - | [ckpt (127 MB)](https://storage.googleapis.com/tf_model_garden/vision/cvrl/r3d_1x_k400_800ep.tar.gz) |
| R3D-50 (1x) | 31.7M  | Kinetics-600 | 800 | 68.5% | 70.4% | [ckpt (127 MB)](https://storage.googleapis.com/tf_model_garden/vision/cvrl/r3d_1x_k600_800ep.tar.gz) |


## Citation

```
@inproceedings{qian2021spatiotemporal,
  title={Spatiotemporal contrastive video representation learning},
  author={Qian, Rui and Meng, Tianjian and Gong, Boqing and Yang, Ming-Hsuan and Wang, Huisheng and Belongie, Serge and Cui, Yin},
  booktitle={CVPR},
  year={2021}
}
```