This is the implementation of our paper "Video Playback Rate Perception for Self-supervised Spatio-Temporal Representation Learning".

• Install

  Our experiments run on Python 3.6.1 and PyTorch 0.4.1. All dependencies can be installed using pip:

  python -m pip install -r requirements.txt

• Data preparation

  We construct experiments on UCF101 and HMDB51 (the split1 of UCF101 for pre-training and the rest for fine-tuning). The expected dataset directory hierarchy is as follow:

  ├── UCF101/HMDB51
  │   ├── split
  │   │   ├── classInd.txt
  │   │   ├── testlist01.txt
  │   │   ├── trainlist01.txt
  │   │   └── ...
  │   └── video
  │       ├── ApplyEyeMakeup
  │       │   └── *.avi
  │       └── ...
  └── ...
  

• Train and Test Pre-training on Pretext Task

  python train_predict.py --gpu 0 --epoch 300 --model_name c3d/r21d/r3d

  Action Recognition

  python ft_classfy.py --gpu 0 --model_name c3d/r21d/r3d --pre_path [your pre-trained model] --split 1/2/3
  python test_classify.py

  Video Retrieval

  Please refer to the code video_retrieval_samples.py of VCOP.

• Models

  Pre-trained PRP model on the split1 of UCF101: C3D(OneDrive); R3D(OneDrive); R(2+1)D(OneDrive)

• Action Recognition Results

  Architecture	UCF101(%)	HMDB51(%)
  C3D	69.1	34.5
  R3D	66.5	29.7
  R(2+1)D	72.1	35.0

This project is released under the Apache 2.0 license.

Please cite the following paper if you feel PRP useful to your research

@InProceedings{Yao_2020_CVPR,  
author = {Yao, Yuan and Liu, Chang and Luo, Dezhao and Zhou, Yu and Ye, Qixiang},  
title = {Video Playback Rate Perception for Self-Supervised Spatio-Temporal Representation Learning},  
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},  
month = {June},  
year = {2020}  
}