Awesome

Semi-Supervised Action Recognition with Temporal Contrastive Learning [Paper] [Website]

This repository contains the implementation details of our Temporal Contrastive Learning (TCL) approach for action recognition in videos.

Ankit Singh*, Omprakash Chakraborty*, Ashutosh Varshney, Rameswar Panda, Rogerio Feris, Kate Saenko and Abir Das, "Semi-Supervised Action Recognition with Temporal Contrastive Learning"
*: Equal contributions

If you use the codes and models from this repo, please cite our work. Thanks!

@InProceedings{Singh_2021_CVPR,
    author    = {Singh, Ankit and Chakraborty, Omprakash and Varshney, Ashutosh and Panda, Rameswar and Feris, Rogerio and Saenko, Kate and Das, Abir},
    title     = {Semi-Supervised Action Recognition With Temporal Contrastive Learning},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {10389-10399}
}

Requirements

The code is written for python 3.6.10, but should work for other version with some modifications.

pip install -r requirements.txt

Data Preparation

The dataloader (ops/dataset.py) can load videos (image sequences) stored in the following format:

-- dataset_dir
---- data
------category.txt  
------train.txt
------val.txt 
---- Frames
------ video_0_folder
-------- 00001.jpg
-------- 00002.jpg
-------- ...
------ video_1_folder
------ ...

For each dataset, root_dataset.yaml should contain the dataset_dir where each dataset is stored

Each line in train.txt and val.txt includes 3 elements and separated by space. Four elements (in order) include (1)relative paths to video_x_folder from dataset_dir, (2) total number of frames, (3) label id (a numeric number).

E.g., a video_x has 300 frames and belong to label 1.

path/to/video_x_folder 300 1

After that, in the ops/dataset_config.py, the location paths of category.txt, Frames, train.txt and val.txt should be included accordingly.

Samples for some datasets are already mentioned in the respective files.

We provided three sample scripts in the tools folder to help convert some datasets but the details in the scripts must be set accordingly. E.g., the path to videos.

Mini-datasets

We provide the category.txt, train.txt and val.txt for the Mini-Something-Something V2 dataset.

Python script overview

main.py - It contains the code for Temporal Contrastive Learning(TCL) with the 2 pathway model.

opts.py - It contains the file with default value for different parameter used in 2 pathway model.

ops/dataset_config.py - It contains the code for different config for different dataset and their location e.g Kinetics, Jester, SomethingV2

ops/dataset.py - It contains the code for how frames are sampled from video

Key Parameters:

use_group_contrastive: to use group contrastive loss
use_finetuning : option to use finetuning at the last
finetune_start_epoch: from which epoch to start finetuning
finetune_lr: if want to use different lr other than normal one
gamma_finetune: weight for pl_loss in finetuning step
finetune_stage_eval_freq: printing freq for finetuning stage
threshold: used in fine tuning step for selection of labels
sup_thresh: till which epoch supervised only to be run
percentage: percentage of unlabeled data e.g 0.99 ,0.95
gamma: weight of instance contrastive loss
lr: starting learning rate
mu: ratio of unlabeled to labeled data
flip: whether to use horizontal flip in transforms or not

Training TCL

For running x% labeled data scenario, it expects to have a folder named Run_x where all the labeled and unlabeled data will be split as per the input seed.
All the models and logs will be stored inside a sub folder of checkpoints directory. A different subfolder will be created on each execution.

Sample Code to train TCL

python main.py somethingv2 RGB --seed 123 --strategy classwise --arch resnet18 --num_segments 8 --second_segments 4 --threshold 0.8 --gd 20 --lr 0.02 --wd 1e-4 --epochs 400 --percentage 0.95 --batch-size 8 -j 16 --dropout 0.5 --consensus_type=avg --eval-freq=1 --print-freq 50 --shift --shift_div=8 --shift_place=blockres --npb --gpus 0 1 --mu 3 --gamma 9 --gamma_finetune 1 --use_group_contrastive --use_finetuning --finetune_start_epoch 350 --sup_thresh 50 --valbatchsize 16 --finetune_lr 0.002

Reference

The implementation reused some portions from TSM[1].

Lin, Ji, Chuang Gan, and Song Han. "Tsm: Temporal shift module for efficient video understanding." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019.