The source code for paper Enhancing Unsupervised Video Representation Learning by Decoupling the Scene and the Motion

1. Introduction (scene-dominated to motion-dominated)


Video datasets are usually scene-dominated, We propose to decouple the scene and the motion (DSM) with two simple operations, so that the model attention towards the motion information is better paid.

The generated triplet is as below:

What DSM learned?

With DSM pretrain, the model learn to focus on motion region (Not necessarily actor) powerful without one label available.


2. Installation


Please refer dataset.md for details.


  • Python3
  • pytorch1.1+
  • PIL
  • Intel (on the fly decode)

3. Structure

  • datasets
    • list
      • hmdb51: the train/val lists of HMDB51
      • ucf101: the train/val lists of UCF101
      • kinetics-400: the train/val lists of kinetics-400
      • diving48: the train/val lists of diving48
  • experiments
    • logs: experiments record in detials
    • gradientes: grad check
    • visualization:
  • src
    • data: load data
    • loss: the loss evaluate in this paper
    • model: network architectures
    • scripts: train/eval scripts
    • augment: detail implementation of Spatio-temporal Augmentation
    • utils
    • feature_extract.py: feature extractor given pretrained model
    • main.py: the main function of finetune
    • trainer.py
    • option.py
    • pt.py: self-supervised pretrain
    • ft.py: supervised finetune


Self-supervised Pretrain

bash scripts/kinetics/pt.sh
bash scripts/ucf101/pt.sh

Supervised Finetune (Clip-level)

bash scripts/hmdb51/ft.sh
bash scripts/ucf101/ft.sh
bash scripts/kinetics/ft.sh

Video-level Evaluation

Following common practice TSN and Non-local.
The final video-level result is average by 10 temporal window sampling + corner crop, which lead to better result than clip-level.
Refer test.py for details.

Pretrain And Eval In one step

bash scripts/hmdb51/pt_and_ft_hmdb51.sh

Notice: More Training Options and ablation study Can be find in scripts

Video Retrieve and other visualization

(1). Feature Extractor

As STCR can be easily extend to other video representation task, we offer the scripts to perform feature extract.

python feature_extractor.py

The feature will be saved as a single numpy file in the format [video_nums,features_dim] for further visualization.

(2). Reterival Evaluation

modify line60-line62 in reterival.py.

python reterival.py


Action Recognition

UCF101 Pretrained (I3D)

Method UCF101 HMDB51
Random Initialization 47.9 29.6
MoCo Baseline 62.3 36.5
DSM(Triplet) 70.7 48.5
DSM 74.8 52.5

Kinetics Pretrained

Video Retrieve (UCF101-C3D)

Method @1 @5 @10 @20 @50
DSM 16.8 33.4 43.4 54.6 70.7

Video Retrieve (HMDB51-C3D)

Method @1 @5 @10 @20 @50
DSM 8.2 25.9 38.1 52.0 75.0

More Visualization


This work is partly based on STN, UEL and MoCo.