The source code for our paper

"Self-Supervised Temporal-Discriminative Representation Learning for Video Action Recognition" paper

Without one label available, our method learn to focus on motion region powerful!

Our self-supervised VTDL signifcantly outperforms existing self-supervised learning method in video action recognition, even achieve better result than fully-supervised methods on UCF101 and HMDB51 when a small-scale video dataset (with only thousands of videos) is used for pre-training!

• Python3
• pytorch1.1+
• PIL

• 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
• experiments
  • logs: experiments record in detials
  • TemporalDis
    • hmdb51
    • ucf101
    • kinetics
  • gradientes:
  • visualization
• src
  • data: load data
  • loss: the loss evluate in this paper
  • model: network architectures
  • scripts: train/eval scripts
  • TC: detail implementation of Spatio-temporal consistency
  • utils
  • feature_extract.py
  • main.py
  • trainer.py
  • option.py

Look dataset.md. Prepare dataset in txt file, and each row of txt is as below: The split of hmdb51/ucf101/kinetics-400 can be download from google driver.

Each item include

video_path class frames_num

The network is in the folder src/model/[backbone].py

bash scripts/TemporalDisc/hmdb51.sh

bash scripts/TemporalDisc/ucf101.sh

bash scripts/TemporalDisc/kinetics.sh

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

#!/usr/bin/env bash
python main.py \
--method ft \
--train_list ../datasets/lists/hmdb51/hmdb51_rgb_train_split_1.txt \
--val_list ../datasets/lists/hmdb51/hmdb51_rgb_val_split_1.txt \
--dataset hmdb51 \
--arch i3d \
--mode rgb \
--lr 0.001 \
--lr_steps 10 20 25 30 35 40 \
--epochs 45 \
--batch_size 4 \
--data_length 64 \
--workers 8 \
--dropout 0.5 \
--gpus 2 \
--logs_path ../experiments/logs/hmdb51_i3d_ft \
--print-freq 100 \
--weights ../experiments/TemporalDis/hmdb51/models/04-16-2328_aug_CJ/ckpt_epoch_48.pth

#!/usr/bin/env bash
python main.py \
--method ft \
--train_list ../datasets/lists/ucf101/ucf101_rgb_train_split_1.txt \
--val_list ../datasets/lists/ucf101/ucf101_rgb_val_split_1.txt \
--dataset ucf101 \
--arch i3d \
--mode rgb \
--lr 0.0005 \
--lr_steps 10 20 25 30 35 40 \
--epochs 45 \
--batch_size 4 \
--data_length 64 \
--workers 8 \
--dropout 0.5 \
--gpus 2 \
--logs_path ../experiments/logs/ucf101_i3d_ft \
--print-freq 100 \
--weights ../experiments/TemporalDis/ucf101/models/04-18-2208_aug_CJ/ckpt_epoch_45.pth

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

With same experiment setting, the result is reported below:

We provided trained models/logs/performance in google driver.

performance;

trained_model;

logs

performance;

trained_model;

logs

(a). Pretrain

Loss curve:

Ins Prob:

pretrained_weight

This pretrained model can achieve 52.7% on HMDB51.

(b). Finetune

performance;

trained_model;

logs

The result is report with single video clip. In the test, we will average ten clips as final predictions. Will lead to around 2-3% improvement.

python test.py

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]

Please cite our paper if you find this code useful for your research.

@Article{wang2020self,
  author  = {Jinpeng Wang and Yiqi Lin and Andy J. Ma and Pong C. Yuen},
  title   = {Self-supervised Temporal Discriminative Learning for Video Representation Learning},
  journal = {arXiv preprint arXiv:2008.02129},
  year    = {2020},
}

The project is partly based on Unsupervised Embedding Learning and MOCO.